|
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 |
|
|
kusano |
2b45e8 |
#define SP r12
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define M r32
|
|
kusano |
2b45e8 |
#define N r33
|
|
kusano |
2b45e8 |
#ifndef XDOUBLE
|
|
kusano |
2b45e8 |
#define A r36
|
|
kusano |
2b45e8 |
#define LDA r37
|
|
kusano |
2b45e8 |
#define X1 r38
|
|
kusano |
2b45e8 |
#define INCX r39
|
|
kusano |
2b45e8 |
#define Y1 r34
|
|
kusano |
2b45e8 |
#define INCY r35
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define A r38
|
|
kusano |
2b45e8 |
#define LDA r39
|
|
kusano |
2b45e8 |
#define X1 r34
|
|
kusano |
2b45e8 |
#define INCX r35
|
|
kusano |
2b45e8 |
#define Y1 r36
|
|
kusano |
2b45e8 |
#define INCY r37
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define BUFFER r11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define I r15
|
|
kusano |
2b45e8 |
#define J r16
|
|
kusano |
2b45e8 |
#define AO1 r17
|
|
kusano |
2b45e8 |
#define AO2 r18
|
|
kusano |
2b45e8 |
#define AO3 r19
|
|
kusano |
2b45e8 |
#define AO4 r20
|
|
kusano |
2b45e8 |
#define AO5 r21
|
|
kusano |
2b45e8 |
#define AO6 r22
|
|
kusano |
2b45e8 |
#define AO7 r23
|
|
kusano |
2b45e8 |
#define AO8 r24
|
|
kusano |
2b45e8 |
#define X2 r25
|
|
kusano |
2b45e8 |
#define Y2 r26
|
|
kusano |
2b45e8 |
#define LDA7M8 r27
|
|
kusano |
2b45e8 |
#define INCX5 r28
|
|
kusano |
2b45e8 |
#define INCY5 r29
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define YY1 r8
|
|
kusano |
2b45e8 |
#define YY2 r9
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ARLC r30
|
|
kusano |
2b45e8 |
#define PR r31
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#ifdef DOUBLE
|
|
kusano |
2b45e8 |
#define RPREFETCH (16 * 3 + 8)
|
|
kusano |
2b45e8 |
#else
|
|
kusano |
2b45e8 |
#define RPREFETCH (16 * 3 + 16)
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
#define PREFETCH lfetch.nt1
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
#define ALPHA f6
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
PROLOGUE
|
|
kusano |
2b45e8 |
.prologue
|
|
kusano |
2b45e8 |
PROFCODE
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
mov ARLC = ar.lc
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
adds r15 = 24, SP
|
|
kusano |
2b45e8 |
adds r14 = 16, SP
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
#ifdef XDOUBLE
|
|
kusano |
2b45e8 |
ld8 X1 = [r14], 16
|
|
kusano |
2b45e8 |
ld8 INCX = [r15], 16
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
#endif
|
|
kusano |
2b45e8 |
ld8 Y1 = [r14], 16
|
|
kusano |
2b45e8 |
ld8 INCY = [r15], 16
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
ld8 BUFFER = [r14]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov PR = pr
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ALPHA = f8
|
|
kusano |
2b45e8 |
.body
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
cmp.ge p7, p0 = r0, M
|
|
kusano |
2b45e8 |
cmp.ge p6, p0 = r0, N
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
shladd INCX = INCX, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
shladd INCY = INCY, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
shladd LDA = LDA, BASE_SHIFT, r0
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mbb
|
|
kusano |
2b45e8 |
(p7) br.cond.dpnt .L999
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L999
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd INCY5 = INCY, 2, INCY
|
|
kusano |
2b45e8 |
shladd INCX5 = INCX, 2, INCX
|
|
kusano |
2b45e8 |
cmp.eq p10, p0 = SIZE, INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p10) mov BUFFER = X1
|
|
kusano |
2b45e8 |
(p10) br.cond.dptk .L10
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mov pr.rot= 0
|
|
kusano |
2b45e8 |
shladd X2 = INCX, 2, X1
|
|
kusano |
2b45e8 |
mov YY1 = BUFFER
|
|
kusano |
2b45e8 |
adds YY2 = 4 * SIZE, BUFFER
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shr I = M, 3
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
mov ar.ec= 5
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mib
|
|
kusano |
2b45e8 |
cmp.gt p6, p0 = 0, I
|
|
kusano |
2b45e8 |
tbit.nz p13, p0 = M, 2
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L05
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L01:
|
|
kusano |
2b45e8 |
(p20) STFD [YY1] = f36, SIZE
|
|
kusano |
2b45e8 |
(p20) STFD [YY2] = f56, SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f32 = [X1], INCX
|
|
kusano |
2b45e8 |
(p16) LDFD f52 = [X2], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p20) STFD [YY1] = f41, SIZE
|
|
kusano |
2b45e8 |
(p20) STFD [YY2] = f61, SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f37 = [X1], INCX
|
|
kusano |
2b45e8 |
(p16) LDFD f57 = [X2], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p20) STFD [YY1] = f46, SIZE
|
|
kusano |
2b45e8 |
(p20) STFD [YY2] = f66, SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f42 = [X1], INCX
|
|
kusano |
2b45e8 |
(p16) LDFD f62 = [X2], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p20) STFD [YY1] = f51, 5 * SIZE
|
|
kusano |
2b45e8 |
(p20) STFD [YY2] = f71, 5 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f47 = [X1], INCX5
|
|
kusano |
2b45e8 |
(p16) LDFD f67 = [X2], INCX5
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L01
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L05:
|
|
kusano |
2b45e8 |
(p13) LDFD f32 = [X1], INCX
|
|
kusano |
2b45e8 |
tbit.nz p14, p0 = M, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f33 = [X1], INCX
|
|
kusano |
2b45e8 |
tbit.nz p15, p0 = M, 0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f34 = [X1], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f35 = [X1], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f36 = [X1], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) STFD [YY1] = f32, SIZE
|
|
kusano |
2b45e8 |
(p14) LDFD f37 = [X1], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) STFD [YY1] = f33, SIZE
|
|
kusano |
2b45e8 |
(p15) LDFD f38 = [X1], INCX
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) STFD [YY1] = f34, SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) STFD [YY1] = f35, SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) STFD [YY1] = f36, SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) STFD [YY1] = f37, SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) STFD [YY1] = f38, SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L10:
|
|
kusano |
2b45e8 |
mov YY1 = Y1
|
|
kusano |
2b45e8 |
shladd Y2 = INCY, 2, Y1
|
|
kusano |
2b45e8 |
shladd YY2 = INCY, 2, Y1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
nop __LINE__
|
|
kusano |
2b45e8 |
shr J = N, 3
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mib
|
|
kusano |
2b45e8 |
nop __LINE__
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = r0, J
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L20
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L11:
|
|
kusano |
2b45e8 |
mov AO1 = A
|
|
kusano |
2b45e8 |
adds AO2 = 1 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO3 = 2 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO4 = 3 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO5 = 4 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO6 = 5 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO7 = 6 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO8 = 7 * SIZE, A
|
|
kusano |
2b45e8 |
shladd A = LDA, 3, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd LDA7M8 = LDA, 3, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
sub LDA7M8 = LDA, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds LDA7M8 = 8 * SIZE, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov f8 = f0
|
|
kusano |
2b45e8 |
mov f9 = f0
|
|
kusano |
2b45e8 |
mov f10 = f0
|
|
kusano |
2b45e8 |
mov f11 = f0
|
|
kusano |
2b45e8 |
mov f12 = f0
|
|
kusano |
2b45e8 |
mov f13 = f0
|
|
kusano |
2b45e8 |
mov f14 = f0
|
|
kusano |
2b45e8 |
mov f15 = f0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mov pr.rot= 0
|
|
kusano |
2b45e8 |
shr I = M, 3
|
|
kusano |
2b45e8 |
mov ar.ec = 2
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov X1 = BUFFER
|
|
kusano |
2b45e8 |
adds X2 = 2 * SIZE, BUFFER
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = -1, I
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L12:
|
|
kusano |
2b45e8 |
(p16) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f39 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f40 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f42 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f44 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f46 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f41 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f43 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f45 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f47 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f48 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f50 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f52 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f54 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f49 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f51 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f53 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f55 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f56 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f58 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f60 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f62 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f57 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f59 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f61 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f63 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f64 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f66 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f68 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f70 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f65 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f67 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f69 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f71 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f72 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f74 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f76 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f78 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f73 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f75 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f77 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f79 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f80 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f82 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f84 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f86 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f81 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f83 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f85 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f87 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f88 = [AO1], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f90 = [AO3], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f92 = [AO5], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f94 = [AO7], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f89 = [AO2], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f91 = [AO4], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f93 = [AO6], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f95 = [AO8], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f102 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f101 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f103 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f96, f48, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f96, f56, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f96, f64, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f96, f72, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f96, f80, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f96, f88, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f97, f49, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f97, f57, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f97, f65, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f97, f73, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f97, f81, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f97, f89, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f98, f50, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f98, f58, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f98, f66, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f98, f74, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f98, f82, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f98, f90, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f99, f51, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f99, f59, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f99, f67, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f99, f75, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f99, f83, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f99, f91, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f100, f52, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f100, f60, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f100, f68, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f100, f76, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f100, f84, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f100, f92, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f101, f53, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f101, f61, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f101, f69, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f101, f77, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f101, f85, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f101, f93, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f102, f54, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f102, f62, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f102, f70, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f102, f78, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f102, f86, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f102, f94, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f103, f39, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f103, f47, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f103, f55, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f103, f63, f11
|
|
kusano |
2b45e8 |
(p16) FMA f12 = f103, f71, f12
|
|
kusano |
2b45e8 |
(p16) FMA f13 = f103, f79, f13
|
|
kusano |
2b45e8 |
(p16) FMA f14 = f103, f87, f14
|
|
kusano |
2b45e8 |
(p16) FMA f15 = f103, f95, f15
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L12
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L15:
|
|
kusano |
2b45e8 |
tbit.nz p13, p11 = M, 2
|
|
kusano |
2b45e8 |
tbit.nz p14, p12 = M, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p11) adds AO5 = - 4 * SIZE, AO5
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mbb
|
|
kusano |
2b45e8 |
(p11) adds AO7 = - 4 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p13) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
tbit.nz p15, p0 = M, 0
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p14) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p11) adds AO6 = - 4 * SIZE, AO6
|
|
kusano |
2b45e8 |
(p12) adds AO7 = - 2 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f40 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f42 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f44 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f46 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f41 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f43 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f45 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f48 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f50 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f52 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f54 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f49 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f51 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f53 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f56 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f58 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f60 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f62 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f57 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f59 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f61 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f64 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f66 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f68 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f70 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f65 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f67 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f69 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f72 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f74 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f76 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f78 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f73 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f75 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f77 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f80 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f82 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f84 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f86 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f81 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f83 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f85 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f88 = [AO1]
|
|
kusano |
2b45e8 |
(p13) LDFD f90 = [AO3]
|
|
kusano |
2b45e8 |
(p14) LDFD f92 = [AO5]
|
|
kusano |
2b45e8 |
(p15) LDFD f94 = [AO7]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f89 = [AO2]
|
|
kusano |
2b45e8 |
(p13) LDFD f91 = [AO4]
|
|
kusano |
2b45e8 |
(p14) LDFD f93 = [AO6]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f101 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) LDFD f102 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f96, f48, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f96, f56, f11
|
|
kusano |
2b45e8 |
(p13) FMA f12 = f96, f64, f12
|
|
kusano |
2b45e8 |
(p13) FMA f13 = f96, f72, f13
|
|
kusano |
2b45e8 |
(p13) FMA f14 = f96, f80, f14
|
|
kusano |
2b45e8 |
(p13) FMA f15 = f96, f88, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f97, f49, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f97, f57, f11
|
|
kusano |
2b45e8 |
(p13) FMA f12 = f97, f65, f12
|
|
kusano |
2b45e8 |
(p13) FMA f13 = f97, f73, f13
|
|
kusano |
2b45e8 |
(p13) FMA f14 = f97, f81, f14
|
|
kusano |
2b45e8 |
(p13) FMA f15 = f97, f89, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f98, f50, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f98, f58, f11
|
|
kusano |
2b45e8 |
(p13) FMA f12 = f98, f66, f12
|
|
kusano |
2b45e8 |
(p13) FMA f13 = f98, f74, f13
|
|
kusano |
2b45e8 |
(p13) FMA f14 = f98, f82, f14
|
|
kusano |
2b45e8 |
(p13) FMA f15 = f98, f90, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f99, f51, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f99, f59, f11
|
|
kusano |
2b45e8 |
(p13) FMA f12 = f99, f67, f12
|
|
kusano |
2b45e8 |
(p13) FMA f13 = f99, f75, f13
|
|
kusano |
2b45e8 |
(p13) FMA f14 = f99, f83, f14
|
|
kusano |
2b45e8 |
(p13) FMA f15 = f99, f91, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
(p14) FMA f10 = f100, f52, f10
|
|
kusano |
2b45e8 |
(p14) FMA f11 = f100, f60, f11
|
|
kusano |
2b45e8 |
(p14) FMA f12 = f100, f68, f12
|
|
kusano |
2b45e8 |
(p14) FMA f13 = f100, f76, f13
|
|
kusano |
2b45e8 |
(p14) FMA f14 = f100, f84, f14
|
|
kusano |
2b45e8 |
(p14) FMA f15 = f100, f92, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
(p14) FMA f10 = f101, f53, f10
|
|
kusano |
2b45e8 |
(p14) FMA f11 = f101, f61, f11
|
|
kusano |
2b45e8 |
(p14) FMA f12 = f101, f69, f12
|
|
kusano |
2b45e8 |
(p14) FMA f13 = f101, f77, f13
|
|
kusano |
2b45e8 |
(p14) FMA f14 = f101, f85, f14
|
|
kusano |
2b45e8 |
(p14) FMA f15 = f101, f93, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p15) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
(p15) FMA f10 = f102, f54, f10
|
|
kusano |
2b45e8 |
(p15) FMA f11 = f102, f62, f11
|
|
kusano |
2b45e8 |
(p15) FMA f12 = f102, f70, f12
|
|
kusano |
2b45e8 |
(p15) FMA f13 = f102, f78, f13
|
|
kusano |
2b45e8 |
(p15) FMA f14 = f102, f86, f14
|
|
kusano |
2b45e8 |
(p15) FMA f15 = f102, f94, f15
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f32 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f33 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f34 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f35 = [Y1], INCY5
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f36 = [Y2], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f37 = [Y2], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f38 = [Y2], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f39 = [Y2], INCY5
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA f32 = ALPHA, f8, f32
|
|
kusano |
2b45e8 |
FMA f33 = ALPHA, f9, f33
|
|
kusano |
2b45e8 |
FMA f34 = ALPHA, f10, f34
|
|
kusano |
2b45e8 |
FMA f35 = ALPHA, f11, f35
|
|
kusano |
2b45e8 |
FMA f36 = ALPHA, f12, f36
|
|
kusano |
2b45e8 |
FMA f37 = ALPHA, f13, f37
|
|
kusano |
2b45e8 |
FMA f38 = ALPHA, f14, f38
|
|
kusano |
2b45e8 |
FMA f39 = ALPHA, f15, f39
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f32
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f33
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f34
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f35
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY5
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY2] = f36
|
|
kusano |
2b45e8 |
add YY2 = YY2, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY2] = f37
|
|
kusano |
2b45e8 |
add YY2 = YY2, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY2] = f38
|
|
kusano |
2b45e8 |
add YY2 = YY2, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY2] = f39
|
|
kusano |
2b45e8 |
add YY2 = YY2, INCY5
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds J = -1, J
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.lt p6, p0 = 0, J
|
|
kusano |
2b45e8 |
(p6) br.cond.dptk .L11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L20:
|
|
kusano |
2b45e8 |
tbit.z p6, p0 = N, 2
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L30
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov AO1 = A
|
|
kusano |
2b45e8 |
adds AO2 = 1 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO3 = 2 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO4 = 3 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO5 = 4 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO6 = 5 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO7 = 6 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO8 = 7 * SIZE, A
|
|
kusano |
2b45e8 |
shladd A = LDA, 2, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd LDA7M8 = LDA, 2, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
sub LDA7M8 = LDA, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds LDA7M8 = 8 * SIZE, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov f8 = f0
|
|
kusano |
2b45e8 |
mov f9 = f0
|
|
kusano |
2b45e8 |
mov f10 = f0
|
|
kusano |
2b45e8 |
mov f11 = f0
|
|
kusano |
2b45e8 |
mov f12 = f0
|
|
kusano |
2b45e8 |
mov f13 = f0
|
|
kusano |
2b45e8 |
mov f14 = f0
|
|
kusano |
2b45e8 |
mov f15 = f0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mov pr.rot= 0
|
|
kusano |
2b45e8 |
shr I = M, 3
|
|
kusano |
2b45e8 |
mov ar.ec = 2
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov X1 = BUFFER
|
|
kusano |
2b45e8 |
adds X2 = 2 * SIZE, BUFFER
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = -1, I
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L25
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L22:
|
|
kusano |
2b45e8 |
(p16) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f39 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f40 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f42 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f44 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f46 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f41 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f43 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f45 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f47 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f48 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f50 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f52 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f54 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f49 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f51 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f53 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f55 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f56 = [AO1], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f58 = [AO3], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f60 = [AO5], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f62 = [AO7], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f57 = [AO2], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f59 = [AO4], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f61 = [AO6], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f63 = [AO8], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f102 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f101 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f103 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f96, f48, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f96, f56, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f97, f49, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f97, f57, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f98, f50, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f98, f58, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f99, f51, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f99, f59, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f100, f52, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f100, f60, f11
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f101, f53, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f101, f61, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f102, f54, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f102, f62, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f103, f39, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f103, f47, f9
|
|
kusano |
2b45e8 |
(p16) FMA f10 = f103, f55, f10
|
|
kusano |
2b45e8 |
(p16) FMA f11 = f103, f63, f11
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L22
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L25:
|
|
kusano |
2b45e8 |
tbit.nz p13, p11 = M, 2
|
|
kusano |
2b45e8 |
tbit.nz p14, p12 = M, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p11) adds AO5 = - 4 * SIZE, AO5
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mbb
|
|
kusano |
2b45e8 |
(p11) adds AO7 = - 4 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p13) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
tbit.nz p15, p0 = M, 0
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p14) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p11) adds AO6 = - 4 * SIZE, AO6
|
|
kusano |
2b45e8 |
(p12) adds AO7 = - 2 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f40 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f42 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f44 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f46 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f41 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f43 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f45 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f48 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f50 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f52 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f54 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f49 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f51 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f53 = [AO6], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f56 = [AO1]
|
|
kusano |
2b45e8 |
(p13) LDFD f58 = [AO3]
|
|
kusano |
2b45e8 |
(p14) LDFD f60 = [AO5]
|
|
kusano |
2b45e8 |
(p15) LDFD f62 = [AO7]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f57 = [AO2]
|
|
kusano |
2b45e8 |
(p13) LDFD f59 = [AO4]
|
|
kusano |
2b45e8 |
(p14) LDFD f61 = [AO6]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f101 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) LDFD f102 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f96, f48, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f96, f56, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f97, f49, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f97, f57, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f98, f50, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f98, f58, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
(p13) FMA f10 = f99, f51, f10
|
|
kusano |
2b45e8 |
(p13) FMA f11 = f99, f59, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
(p14) FMA f10 = f100, f52, f10
|
|
kusano |
2b45e8 |
(p14) FMA f11 = f100, f60, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
(p14) FMA f10 = f101, f53, f10
|
|
kusano |
2b45e8 |
(p14) FMA f11 = f101, f61, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p15) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
(p15) FMA f10 = f102, f54, f10
|
|
kusano |
2b45e8 |
(p15) FMA f11 = f102, f62, f11
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f32 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f33 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f34 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f35 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA f32 = ALPHA, f8, f32
|
|
kusano |
2b45e8 |
FMA f33 = ALPHA, f9, f33
|
|
kusano |
2b45e8 |
FMA f34 = ALPHA, f10, f34
|
|
kusano |
2b45e8 |
FMA f35 = ALPHA, f11, f35
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f32
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f33
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f34
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f35
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L30:
|
|
kusano |
2b45e8 |
tbit.z p6, p0 = N, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L40
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov AO1 = A
|
|
kusano |
2b45e8 |
adds AO2 = 1 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO3 = 2 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO4 = 3 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO5 = 4 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO6 = 5 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO7 = 6 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO8 = 7 * SIZE, A
|
|
kusano |
2b45e8 |
shladd A = LDA, 1, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
shladd LDA7M8 = LDA, 1, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
sub LDA7M8 = LDA, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds LDA7M8 = 8 * SIZE, LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov f8 = f0
|
|
kusano |
2b45e8 |
mov f9 = f0
|
|
kusano |
2b45e8 |
mov f10 = f0
|
|
kusano |
2b45e8 |
mov f11 = f0
|
|
kusano |
2b45e8 |
mov f12 = f0
|
|
kusano |
2b45e8 |
mov f13 = f0
|
|
kusano |
2b45e8 |
mov f14 = f0
|
|
kusano |
2b45e8 |
mov f15 = f0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mov pr.rot= 0
|
|
kusano |
2b45e8 |
shr I = M, 3
|
|
kusano |
2b45e8 |
mov ar.ec = 2
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov X1 = BUFFER
|
|
kusano |
2b45e8 |
adds X2 = 2 * SIZE, BUFFER
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = -1, I
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L35
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L32:
|
|
kusano |
2b45e8 |
(p16) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p16) LDFD f39 = [AO8], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f40 = [AO1], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f42 = [AO3], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f44 = [AO5], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f46 = [AO7], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f41 = [AO2], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f43 = [AO4], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f45 = [AO6], LDA7M8
|
|
kusano |
2b45e8 |
(p16) LDFD f47 = [AO8], LDA7M8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f102 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f101 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f103 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f103, f39, f8
|
|
kusano |
2b45e8 |
(p16) FMA f9 = f103, f47, f9
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L32
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L35:
|
|
kusano |
2b45e8 |
tbit.nz p13, p11 = M, 2
|
|
kusano |
2b45e8 |
tbit.nz p14, p12 = M, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p11) adds AO5 = - 4 * SIZE, AO5
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mbb
|
|
kusano |
2b45e8 |
(p11) adds AO7 = - 4 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p13) LDFD f32 = [AO1], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f34 = [AO3], LDA
|
|
kusano |
2b45e8 |
tbit.nz p15, p0 = M, 0
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p14) LDFD f36 = [AO5], LDA
|
|
kusano |
2b45e8 |
(p11) adds AO6 = - 4 * SIZE, AO6
|
|
kusano |
2b45e8 |
(p12) adds AO7 = - 2 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f33 = [AO2], LDA
|
|
kusano |
2b45e8 |
(p13) LDFD f35 = [AO4], LDA
|
|
kusano |
2b45e8 |
(p14) LDFD f37 = [AO6], LDA
|
|
kusano |
2b45e8 |
(p15) LDFD f38 = [AO7], LDA
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f40 = [AO1]
|
|
kusano |
2b45e8 |
(p13) LDFD f42 = [AO3]
|
|
kusano |
2b45e8 |
(p14) LDFD f44 = [AO5]
|
|
kusano |
2b45e8 |
(p15) LDFD f46 = [AO7]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f41 = [AO2]
|
|
kusano |
2b45e8 |
(p13) LDFD f43 = [AO4]
|
|
kusano |
2b45e8 |
(p14) LDFD f45 = [AO6]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f101 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) LDFD f102 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f96, f40, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f97, f41, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f98, f42, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
(p13) FMA f9 = f99, f43, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f100, f44, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
(p14) FMA f9 = f101, f45, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
(p15) FMA f9 = f102, f46, f9
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f32 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f33 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA f32 = ALPHA, f8, f32
|
|
kusano |
2b45e8 |
FMA f33 = ALPHA, f9, f33
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f32
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f33
|
|
kusano |
2b45e8 |
add YY1 = YY1, INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L40:
|
|
kusano |
2b45e8 |
tbit.z p6, p0 = N, 0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L999
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov AO1 = A
|
|
kusano |
2b45e8 |
adds AO2 = 1 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO3 = 2 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO4 = 3 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO5 = 4 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO6 = 5 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO7 = 6 * SIZE, A
|
|
kusano |
2b45e8 |
adds AO8 = 7 * SIZE, A
|
|
kusano |
2b45e8 |
add A = LDA, A
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov f8 = f0
|
|
kusano |
2b45e8 |
mov f9 = f0
|
|
kusano |
2b45e8 |
mov f10 = f0
|
|
kusano |
2b45e8 |
mov f11 = f0
|
|
kusano |
2b45e8 |
mov f12 = f0
|
|
kusano |
2b45e8 |
mov f13 = f0
|
|
kusano |
2b45e8 |
mov f14 = f0
|
|
kusano |
2b45e8 |
mov f15 = f0
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
mov pr.rot= 0
|
|
kusano |
2b45e8 |
shr I = M, 3
|
|
kusano |
2b45e8 |
mov ar.ec = 2
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov X1 = BUFFER
|
|
kusano |
2b45e8 |
adds X2 = 2 * SIZE, BUFFER
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
cmp.eq p16, p0 = r0, r0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
adds I = -1, I
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
mov ar.lc = I
|
|
kusano |
2b45e8 |
cmp.eq p6, p0 = -1, I
|
|
kusano |
2b45e8 |
(p6) br.cond.dpnt .L45
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L42:
|
|
kusano |
2b45e8 |
(p16) LDFD f32 = [AO1], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f34 = [AO3], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f36 = [AO5], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f38 = [AO7], 8 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f33 = [AO2], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f35 = [AO4], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f37 = [AO6], 8 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f39 = [AO8], 8 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f102 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) LDFD f101 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p16) LDFD f103 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p16) FMA f8 = f103, f39, f8
|
|
kusano |
2b45e8 |
br.ctop.sptk.few .L42
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L45:
|
|
kusano |
2b45e8 |
tbit.nz p13, p11 = M, 2
|
|
kusano |
2b45e8 |
tbit.nz p14, p12 = M, 1
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p11) adds AO5 = - 4 * SIZE, AO5
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mbb
|
|
kusano |
2b45e8 |
(p11) adds AO7 = - 4 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p13) LDFD f32 = [AO1]
|
|
kusano |
2b45e8 |
(p13) LDFD f34 = [AO3]
|
|
kusano |
2b45e8 |
tbit.nz p15, p0 = M, 0
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
{ .mmi
|
|
kusano |
2b45e8 |
(p14) LDFD f36 = [AO5]
|
|
kusano |
2b45e8 |
(p11) adds AO6 = - 4 * SIZE, AO6
|
|
kusano |
2b45e8 |
(p12) adds AO7 = - 2 * SIZE, AO7
|
|
kusano |
2b45e8 |
}
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f33 = [AO2]
|
|
kusano |
2b45e8 |
(p13) LDFD f35 = [AO4]
|
|
kusano |
2b45e8 |
(p14) LDFD f37 = [AO6]
|
|
kusano |
2b45e8 |
(p15) LDFD f38 = [AO7]
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f96 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f98 = [X2], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) LDFD f97 = [X1], 3 * SIZE
|
|
kusano |
2b45e8 |
(p13) LDFD f99 = [X2], 3 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f100 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) LDFD f101 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) LDFD f102 = [X1], 1 * SIZE
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f96, f32, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f97, f33, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f98, f34, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p13) FMA f8 = f99, f35, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f100, f36, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p14) FMA f8 = f101, f37, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
(p15) FMA f8 = f102, f38, f8
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
LDFD f32 = [Y1], INCY
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
FMA f32 = ALPHA, f8, f32
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
STFD [YY1] = f32
|
|
kusano |
2b45e8 |
.align 16
|
|
kusano |
2b45e8 |
|
|
kusano |
2b45e8 |
.L999:
|
|
kusano |
2b45e8 |
mov ar.lc = ARLC
|
|
kusano |
2b45e8 |
mov pr = PR, -1
|
|
kusano |
2b45e8 |
br.ret.sptk.many b0
|
|
kusano |
2b45e8 |
;;
|
|
kusano |
2b45e8 |
EPILOGUE
|