339 lines
6.9 KiB
ArmAsm
339 lines
6.9 KiB
ArmAsm
#define ASSEMBLER
|
|
|
|
#include "common.h"
|
|
#define N $r4
|
|
#define XX $r5
|
|
#define YY $r6
|
|
#define ALPHA $f0
|
|
#define X $r7
|
|
#define INCX $r8
|
|
#define Y $r9
|
|
#define INCY $r10
|
|
|
|
#define I $r12
|
|
#define TEMP $r13
|
|
#define t1 $r14
|
|
#define t2 $r16
|
|
#define t3 $r15
|
|
#define t4 $r17
|
|
#define a1 $f12
|
|
#define a2 $f13
|
|
#define a3 $f14
|
|
#define a4 $f15
|
|
#define b1 $f16
|
|
#define b2 $f17
|
|
#define b3 $f18
|
|
#define b4 $f19
|
|
#define VX0 $vr8
|
|
#define VX1 $vr20
|
|
#define VX2 $vr21
|
|
#define VX3 $vr22
|
|
#define VXA $vr23
|
|
|
|
PROLOGUE
|
|
|
|
bge $r0, N, .L999
|
|
li.d TEMP, 1
|
|
movgr2fr.d a1, $r0
|
|
ffint.s.l a1, a1
|
|
movgr2fr.d a2, TEMP
|
|
ffint.s.l a2, a2
|
|
fcmp.ceq.s $fcc0, ALPHA, a1
|
|
bcnez $fcc0, .L999
|
|
slli.d TEMP, TEMP, BASE_SHIFT
|
|
slli.d INCX, INCX, BASE_SHIFT
|
|
slli.d INCY, INCY, BASE_SHIFT
|
|
movfr2gr.s t1, ALPHA
|
|
vreplgr2vr.w VXA, t1
|
|
|
|
srai.d I, N, 3
|
|
bne INCX, TEMP, .L20
|
|
bne INCY, TEMP, .L12 // INCX==1 and INCY!=1
|
|
b .L11 // INCX==1 and INCY==1
|
|
.L20:
|
|
bne INCY, TEMP, .L22 // INCX!=1 and INCY!=1
|
|
b .L21 // INCX!=1 and INCY==1
|
|
|
|
.L11:
|
|
bge $r0, I, .L113
|
|
fcmp.ceq.s $fcc0, ALPHA, a2
|
|
bceqz $fcc0, .L112
|
|
.align 3
|
|
|
|
.L111:
|
|
vld VX0, X, 0 * SIZE
|
|
vld VX2, Y, 0 * SIZE
|
|
vld VX1, X, 4 * SIZE
|
|
vld VX3, Y, 4 * SIZE
|
|
vfadd.s VX2, VX0, VX2
|
|
vfadd.s VX3, VX1, VX3
|
|
vst VX2, Y, 0 * SIZE
|
|
vst VX3, Y, 4 * SIZE
|
|
addi.d X, X, 8 * SIZE
|
|
addi.d Y, Y, 8 * SIZE
|
|
addi.d I, I, -1
|
|
blt $r0, I, .L111
|
|
b .L113
|
|
.align 3
|
|
|
|
.L112:
|
|
vld VX0, X, 0 * SIZE
|
|
vld VX2, Y, 0 * SIZE
|
|
vld VX1, X, 4 * SIZE
|
|
vld VX3, Y, 4 * SIZE
|
|
vfmadd.s VX2, VX0, VXA, VX2
|
|
vfmadd.s VX3, VX1, VXA, VX3
|
|
vst VX2, Y, 0 * SIZE
|
|
vst VX3, Y, 4 * SIZE
|
|
addi.d X, X, 8 * SIZE
|
|
addi.d Y, Y, 8 * SIZE
|
|
addi.d I, I, -1
|
|
blt $r0, I, .L112
|
|
b .L113
|
|
.align 3
|
|
|
|
.L113:
|
|
andi I, N, 7
|
|
bge $r0, I, .L999
|
|
.align 3
|
|
|
|
.L114:
|
|
fld.s $f12, X, 0 * SIZE
|
|
fld.s $f14, Y, 0 * SIZE
|
|
addi.d I, I, -1
|
|
fmadd.s $f14, $f12, $f0, $f14
|
|
fst.s $f14, Y, 0 * SIZE
|
|
addi.d X, X, SIZE
|
|
addi.d Y, Y, SIZE
|
|
blt $r0, I, .L114
|
|
b .L999
|
|
.align 3
|
|
|
|
.L12: // INCX==1 and INCY!=1
|
|
bge $r0, I, .L122
|
|
move YY, Y
|
|
.align 3
|
|
|
|
.L121:
|
|
vld VX0, X, 0 * SIZE
|
|
ld.w t1, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t2, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t3, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t4, Y, 0 * SIZE
|
|
vinsgr2vr.w VX2, t1, 0
|
|
vinsgr2vr.w VX2, t2, 1
|
|
vinsgr2vr.w VX2, t3, 2
|
|
vinsgr2vr.w VX2, t4, 3
|
|
add.d Y, Y, INCY
|
|
vfmadd.s VX2, VX0, VXA, VX2
|
|
vld VX1, X, 4 * SIZE
|
|
vstelm.w VX2, YY, 0, 0
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 1
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 2
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 3
|
|
add.d YY, YY, INCY
|
|
ld.w t1, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t2, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t3, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t4, Y, 0 * SIZE
|
|
vinsgr2vr.w VX3, t1, 0
|
|
vinsgr2vr.w VX3, t2, 1
|
|
vinsgr2vr.w VX3, t3, 2
|
|
vinsgr2vr.w VX3, t4, 3
|
|
add.d Y, Y, INCY
|
|
vfmadd.s VX3, VX1, VXA, VX3
|
|
addi.d I, I, -1
|
|
vstelm.w VX3, YY, 0, 0
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 1
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 2
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 3
|
|
add.d YY, YY, INCY
|
|
addi.d X, X, 8 * SIZE
|
|
blt $r0, I, .L121
|
|
.align 3
|
|
|
|
.L122:
|
|
andi I, N, 7
|
|
bge $r0, I, .L999
|
|
.align 3
|
|
|
|
.L123:
|
|
fld.s $f12, X, 0 * SIZE
|
|
fld.s $f14, Y, 0 * SIZE
|
|
addi.d I, I, -1
|
|
fmadd.s $f14, $f12, $f0, $f14
|
|
fst.s $f14, Y, 0 * SIZE
|
|
addi.d X, X, SIZE
|
|
add.d Y, Y, INCY
|
|
blt $r0, I, .L123
|
|
b .L999
|
|
.align 3
|
|
|
|
.L21:// INCX!=1 and INCY==1
|
|
bge $r0, I, .L212
|
|
.align 3
|
|
|
|
.L211:
|
|
vld VX2, Y, 0 * SIZE
|
|
ld.w t1, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t2, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t3, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t4, X, 0 * SIZE
|
|
vinsgr2vr.w VX0, t1, 0
|
|
vinsgr2vr.w VX0, t2, 1
|
|
vinsgr2vr.w VX0, t3, 2
|
|
vinsgr2vr.w VX0, t4, 3
|
|
add.d X, X, INCX
|
|
vfmadd.s VX2, VX0, VXA, VX2
|
|
vld VX3, Y, 4 * SIZE
|
|
vst VX2, Y, 0 * SIZE
|
|
ld.w t1, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t2, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t3, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t4, X, 0 * SIZE
|
|
vinsgr2vr.w VX1, t1, 0
|
|
vinsgr2vr.w VX1, t2, 1
|
|
vinsgr2vr.w VX1, t3, 2
|
|
vinsgr2vr.w VX1, t4, 3
|
|
add.d X, X, INCX
|
|
vfmadd.s VX3, VX1, VXA, VX3
|
|
addi.d I, I, -1
|
|
vst VX3, Y, 4 * SIZE
|
|
addi.d Y, Y, 8 * SIZE
|
|
blt $r0, I, .L211
|
|
.align 3
|
|
|
|
.L212:
|
|
andi I, N, 7
|
|
bge $r0, I, .L999
|
|
.align 3
|
|
|
|
.L213:
|
|
fld.s $f12, X, 0 * SIZE
|
|
fld.s $f14, Y, 0 * SIZE
|
|
addi.d I, I, -1
|
|
fmadd.s $f14, $f12, $f0, $f14
|
|
fst.s $f14, Y, 0 * SIZE
|
|
add.d X, X, INCX
|
|
addi.d Y, Y, SIZE
|
|
blt $r0, I, .L213
|
|
b .L999
|
|
.align 3
|
|
|
|
.L22:
|
|
bge $r0, I, .L223
|
|
move YY, Y
|
|
.align 3
|
|
|
|
.L222:
|
|
ld.w t1, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t2, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t3, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t4, X, 0 * SIZE
|
|
vinsgr2vr.w VX0, t1, 0
|
|
vinsgr2vr.w VX0, t2, 1
|
|
vinsgr2vr.w VX0, t3, 2
|
|
vinsgr2vr.w VX0, t4, 3
|
|
add.d X, X, INCX
|
|
ld.w t1, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t2, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t3, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t4, Y, 0 * SIZE
|
|
vinsgr2vr.w VX2, t1, 0
|
|
vinsgr2vr.w VX2, t2, 1
|
|
vinsgr2vr.w VX2, t3, 2
|
|
vinsgr2vr.w VX2, t4, 3
|
|
add.d Y, Y, INCY
|
|
vfmadd.s VX2, VX0, VXA, VX2
|
|
ld.w t1, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t2, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t3, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
ld.w t4, X, 0 * SIZE
|
|
add.d X, X, INCX
|
|
vinsgr2vr.w VX1, t1, 0
|
|
vinsgr2vr.w VX1, t2, 1
|
|
vinsgr2vr.w VX1, t3, 2
|
|
vinsgr2vr.w VX1, t4, 3
|
|
vstelm.w VX2, YY, 0, 0
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 1
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 2
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX2, YY, 0, 3
|
|
add.d YY, YY, INCY
|
|
ld.w t1, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t2, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t3, Y, 0 * SIZE
|
|
add.d Y, Y, INCY
|
|
ld.w t4, Y, 0 * SIZE
|
|
vinsgr2vr.w VX3, t1, 0
|
|
vinsgr2vr.w VX3, t2, 1
|
|
vinsgr2vr.w VX3, t3, 2
|
|
vinsgr2vr.w VX3, t4, 3
|
|
add.d Y, Y, INCY
|
|
vfmadd.s VX3, VX1, VXA, VX3
|
|
addi.d I, I, -1
|
|
vstelm.w VX3, YY, 0, 0
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 1
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 2
|
|
add.d YY, YY, INCY
|
|
vstelm.w VX3, YY, 0, 3
|
|
add.d YY, YY, INCY
|
|
blt $r0, I, .L222
|
|
.align 3
|
|
|
|
.L223:
|
|
andi I, N, 7
|
|
bge $r0, I, .L999
|
|
.align 3
|
|
|
|
.L224:
|
|
fld.s $f12, X, 0 * SIZE
|
|
fld.s $f14, Y, 0 * SIZE
|
|
addi.d I, I, -1
|
|
fmadd.s $f14, $f12, $f0, $f14
|
|
fst.s $f14, Y, 0 * SIZE
|
|
add.d X, X, INCX
|
|
add.d Y, Y, INCY
|
|
blt $r0, I, .L224
|
|
.align 3
|
|
|
|
.L999:
|
|
move $r4, $r12
|
|
jirl $r0, $r1, 0x0
|
|
.align 3
|
|
|
|
EPILOGUE
|