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

loongarch64: Add and refine iamax optimization functions.

tags/v0.3.26
Hao Chen Shiyou Yin 1 year ago
parent
ea70e165c7
commit
173a65d4e6
12 changed files with 2101 additions and 866 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +4
    -0
      common_loongarch64.h
  2. +5
    -2
      kernel/loongarch64/KERNEL.LOONGSON2K1000
  3. +5
    -2
      kernel/loongarch64/KERNEL.LOONGSON3R5
  4. +194
    -0
      kernel/loongarch64/camax_lasx.S
  5. +206
    -0
      kernel/loongarch64/camax_lsx.S
  6. +209
    -45
      kernel/loongarch64/iamax_lasx.S
  7. +482
    -0
      kernel/loongarch64/iamax_lsx.S
  8. +562
    -0
      kernel/loongarch64/icamax_lasx.S
  9. +434
    -0
      kernel/loongarch64/icamax_lsx.S
  10. +0
    -275
      kernel/loongarch64/idamax_lasx.S
  11. +0
    -267
      kernel/loongarch64/idamax_lsx.S
  12. +0
    -275
      kernel/loongarch64/isamax_lsx.S

+ 4
- 0
common_loongarch64.h View File

@@ -139,6 +139,7 @@ static inline int WhereAmI(void){
#define XVFMAX xvfmax.d
#define XVFMAXA xvfmaxa.d
#define XVCMPEQ xvfcmp.ceq.d
#define XVCMPLE xvfcmp.cle.d
#define XVCMPLT xvfcmp.clt.d
#define XVMUL xvfmul.d
#define XVMSUB xvfmsub.d
@@ -151,6 +152,7 @@ static inline int WhereAmI(void){
#define VFMAX vfmax.d
#define VFMAXA vfmaxa.d
#define VCMPEQ vfcmp.ceq.d
#define VCMPLE vfcmp.cle.d
#define VCMPLT vfcmp.clt.d
#define VMUL vfmul.d
#define VMSUB vfmsub.d
@@ -189,6 +191,7 @@ static inline int WhereAmI(void){
#define XVFMAX xvfmax.s
#define XVFMAXA xvfmaxa.s
#define XVCMPEQ xvfcmp.ceq.s
#define XVCMPLE xvfcmp.cle.s
#define XVCMPLT xvfcmp.clt.s
#define XVMUL xvfmul.s
#define XVMSUB xvfmsub.s
@@ -201,6 +204,7 @@ static inline int WhereAmI(void){
#define VFMAX vfmax.s
#define VFMAXA vfmaxa.s
#define VCMPEQ vfcmp.ceq.s
#define VCMPLE vfcmp.cle.s
#define VCMPLT vfcmp.clt.s
#define VMUL vfmul.s
#define VMSUB vfmsub.s


+ 5
- 2
kernel/loongarch64/KERNEL.LOONGSON2K1000 View File

@@ -9,6 +9,7 @@ DSCALKERNEL = dscal_lsx.S

SAMAXKERNEL = amax_lsx.S
DAMAXKERNEL = amax_lsx.S
CAMAXKERNEL = camax_lsx.S

SAMINKERNEL = amin_lsx.S
DAMINKERNEL = amin_lsx.S
@@ -25,8 +26,10 @@ IDMAXKERNEL = imax_lsx.S
ISMINKERNEL = imin_lsx.S
IDMINKERNEL = imin_lsx.S

ISAMAXKERNEL = isamax_lsx.S
IDAMAXKERNEL = idamax_lsx.S
ISAMAXKERNEL = iamax_lsx.S
IDAMAXKERNEL = iamax_lsx.S
ICAMAXKERNEL = icamax_lsx.S
IZAMAXKERNEL = icamax_lsx.S

ISAMINKERNEL = iamin_lsx.S
IDAMINKERNEL = iamin_lsx.S


+ 5
- 2
kernel/loongarch64/KERNEL.LOONGSON3R5 View File

@@ -9,6 +9,7 @@ DSCALKERNEL = dscal_lasx.S

SAMAXKERNEL = amax_lasx.S
DAMAXKERNEL = amax_lasx.S
CAMAXKERNEL = camax_lasx.S

SAMINKERNEL = amin_lasx.S
DAMINKERNEL = amin_lasx.S
@@ -25,8 +26,10 @@ IDMAXKERNEL = imax_lasx.S
ISMINKERNEL = imin_lasx.S
IDMINKERNEL = imin_lasx.S

ISAMAXKERNEL = isamax_lasx.S
IDAMAXKERNEL = idamax_lasx.S
ISAMAXKERNEL = iamax_lasx.S
IDAMAXKERNEL = iamax_lasx.S
ICAMAXKERNEL = icamax_lasx.S
IZAMAXKERNEL = icamax_lasx.S

ISAMINKERNEL = iamin_lasx.S
IDAMINKERNEL = iamin_lasx.S


+ 194
- 0
kernel/loongarch64/camax_lasx.S View File

@@ -0,0 +1,194 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $f14
#define t2 $f18
#define t3 $f15
#define t4 $f17
#define s1 $f22
#define s2 $f9
#define s3 $f10
#define s4 $f11
#define TEMP $r16
#define a0 $f20
#define a1 $f21
#define x1 $xr9
#define x2 $xr10
#define x3 $xr11
#define x4 $xr12
#define VT0 $xr13
#define VT1 $xr14
#define res0 $xr18
#define neg1 $xr19
#define VX0 $xr20
#define VX1 $xr21
#define VM0 $xr22
#define VM1 $xr23

PROLOGUE
xvxor.v VM0, VM0, VM0
xvxor.v res0, res0, res0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
li.w I, -1
slli.d TEMP, TEMP, ZBASE_SHIFT
slli.d INCX, INCX, ZBASE_SHIFT
xvreplgr2vr.w neg1, I
xvffint.s.w neg1, neg1
srai.d I, N, 3
bne INCX, TEMP, .L20
bge $r0, I, .L23
.align 3

.L10:
xvld VX0, X, 0 * SIZE
xvld VX1, X, 8 * SIZE
addi.d I, I, -1
xvpickev.w x1, VX1, VX0
xvpickod.w x2, VX1, VX0
xvfmul.s x3, neg1, x1
xvfmul.s x4, neg1, x2
xvfcmp.clt.s VT0, x1, res0
xvfcmp.clt.s VT1, x2, res0
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VT1
addi.d X, X, 16 * SIZE
xvfadd.s VM1, x1, x2
xvfmax.s VM0, VM0, VM1
blt $r0, I, .L10
.align 3

.L11:
xvpickve.w x1, VM0, 0
xvpickve.w x2, VM0, 1
xvpickve.w x3, VM0, 2
xvpickve.w x4, VM0, 3
xvfmax.s VM1, x1, x2
xvfmax.s VM0, x3, x4
xvfmax.s VM0, VM0, VM1
b .L23
.align 3

.L20: // INCX!=1
bge $r0, I, .L23
.align 3

.L21:
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
addi.d I, I, -1
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s3, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s4, t1, t3
blt $r0, I, .L21
.align 3

.L22:
fmax.s s1, s1, s2
fmax.s s3, s3, s4
fmax.s s1, s1, s3
.align 3

.L23: //N<8
andi I, N, 7
bge $r0, I, .L999
.align 3

.L24:
LD a0, X, 0 * SIZE
LD a1, X, 1 * SIZE
addi.d I, I, -1
FABS a0, a0
FABS a1, a1
ADD a0, a0, a1
add.d X, X, INCX
fmax.s s1, a0, s1
blt $r0, I, .L24
.align 3

.L999:
fmov.s $f0, $f22
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 206
- 0
kernel/loongarch64/camax_lsx.S View File

@@ -0,0 +1,206 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $f14
#define t2 $f18
#define t3 $f15
#define t4 $f17
#define s1 $f22
#define s2 $f9
#define s3 $f10
#define s4 $f11
#define TEMP $r16
#define a0 $f20
#define a1 $f21
#define x1 $vr9
#define x2 $vr10
#define x3 $vr11
#define x4 $vr12
#define VT0 $vr13
#define VT1 $vr14
#define res0 $vr18
#define neg1 $vr19
#define VX0 $vr20
#define VX1 $vr21
#define VM0 $vr22
#define VM1 $vr23

PROLOGUE
vxor.v VM0, VM0, VM0
vxor.v res0, res0, res0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
li.w I, -1
slli.d TEMP, TEMP, ZBASE_SHIFT
slli.d INCX, INCX, ZBASE_SHIFT
vreplgr2vr.w neg1, I
vffint.s.w neg1, neg1
srai.d I, N, 3
bne INCX, TEMP, .L20
bge $r0, I, .L23
.align 3

.L10:
vld VX0, X, 0 * SIZE
vld VX1, X, 4 * SIZE
addi.d I, I, -1
vpickev.w x1, VX1, VX0
vpickod.w x2, VX1, VX0
vfmul.s x3, neg1, x1
vfmul.s x4, neg1, x2
vfcmp.clt.s VT0, x1, res0
vfcmp.clt.s VT1, x2, res0
vld VX0, X, 8 * SIZE
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VT1
vld VX1, X, 12 * SIZE
vfadd.s VM1, x1, x2
vpickev.w x1, VX1, VX0
vpickod.w x2, VX1, VX0
vfmul.s x3, neg1, x1
vfmul.s x4, neg1, x2
vfcmp.clt.s VT0, x1, res0
vfcmp.clt.s VT1, x2, res0
addi.d X, X, 16 * SIZE
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VT1
vfadd.s x1, x1, x2
vfmax.s VM1, x1, VM1
vfmax.s VM0, VM0, VM1
blt $r0, I, .L10
.align 3

.L11:
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmax.s VM1, x1, x2
vfmax.s VM0, x3, x4
vfmax.s VM0, VM0, VM1
b .L23
.align 3

.L20: // INCX!=1
bge $r0, I, .L23
.align 3

.L21:
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s1, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
addi.d I, I, -1
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s3, t1, t3
fld.s t1, X, 0 * SIZE
fld.s t2, X, 1 * SIZE
add.d X, X, INCX
fld.s t3, X, 0 * SIZE
fld.s t4, X, 1 * SIZE
add.d X, X, INCX
fabs.s t1, t1
fabs.s t2, t2
fabs.s t3, t3
fabs.s t4, t4
fadd.s t1, t1, t2
fadd.s t3, t3, t4
fmax.s s4, t1, t3
blt $r0, I, .L21
.align 3

.L22:
fmax.s s1, s1, s2
fmax.s s3, s3, s4
fmax.s s1, s1, s3
.align 3

.L23: //N<8
andi I, N, 7
bge $r0, I, .L999
.align 3

.L24:
fld.s a0, X, 0 * SIZE
fld.s a1, X, 1 * SIZE
addi.d I, I, -1
fabs.s a0, a0
fabs.s a1, a1
fadd.s a0, a0, a1
add.d X, X, INCX
fmax.s s1, a0, s1
blt $r0, I, .L24
.align 3

.L999:
fmov.s $f0, $f22
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

kernel/loongarch64/isamax_lasx.S → kernel/loongarch64/iamax_lasx.S View File

@@ -1,3 +1,30 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER

#include "common.h"
@@ -39,6 +66,31 @@
slli.d INCX, INCX, BASE_SHIFT
bne INCX, TEMP, .L20
xvld VM0, X, 0
#ifdef DOUBLE
addi.d i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
slli.d i0, i0, 1 //8
xvreplgr2vr.d VINC8, i0
addi.d i0, i0, -15
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 5
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 1 //2
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 2 //3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 3 //4
#else
addi.w i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
@@ -76,9 +128,25 @@
xvinsgr2vr.w VI0, i0, 6 //7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 7 //8
#endif
.align 3

.L10:
#ifdef DOUBLE
xvld VX0, X, 0 * SIZE
xvadd.d VI1, VI1, VINC8
xvld VX1, X, 4 * SIZE
xvadd.d VI2, VI1, VINC4
xvfmaxa.d VM1, VX0, VX1
xvfcmp.ceq.d VT0, VX0, VM1
addi.d I, I, -1
xvbitsel.v VI2, VI2, VI1, VT0
xvfmaxa.d VM1, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
addi.d X, X, 8 * SIZE
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI2, VI0, VT0
#else
xvld VX0, X, 0 * SIZE
addi.d I, I, -1
xvadd.w VI1, VI1, VINC8
@@ -87,10 +155,21 @@
addi.d X, X, 8 * SIZE
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
#endif
blt $r0, I, .L10
.align 3

.L15:
#ifdef DOUBLE
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
#else
xvxor.v VX0, VX0, VX0
xvor.v VX0, VI0, VX0
xvxor.v VX1, VX1, VX1
@@ -103,28 +182,62 @@
xvpickve.w x2, VM0, 1
xvpickve.w x3, VM0, 2
xvpickve.w x4, VM0, 3
xvfmaxa.s VM1, x1, x2
xvfcmp.ceq.s VT0, x1, VM1
#endif
XVFMAXA VM1, x1, x2
XVCMPEQ VT0, x1, VM1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmaxa.s VM0, x3, x4
xvfcmp.ceq.s VT0, x3, VM0
XVFMAXA VM0, x3, x4
XVCMPEQ VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmaxa.s VM0, VM0, VM1
xvfcmp.ceq.s VT0, VM0, VM1
XVFMAXA VM0, VM0, VM1
XVCMPEQ VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.w $f17, TEMP
ffint.s.w $f17, $f17
xvfcmp.ceq.s VT0, VM0, x1
fcmp.ceq.s $fcc0, $f23, $f17
CMPEQ $fcc0, $f15, $f9
bceqz $fcc0, .L26
xvfcmp.clt.s VT0, VI1, VI0
XVCMPLT VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
b .L26
.align 3

.L20: // INCX!=1
move TEMP, X
#ifdef DOUBLE
addi.d i0, i0, 1
ld.d t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
xvinsgr2vr.d VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.d t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.d t3, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.d t4, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
xvinsgr2vr.d VM0, t2, 1
xvinsgr2vr.d VM0, t3, 2
xvinsgr2vr.d VM0, t4, 3
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
slli.d i0, i0, 1 //8
xvreplgr2vr.d VINC8, i0
addi.d i0, i0, -15
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 5
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 1 //2
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 2 //3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 3 //4
#else
addi.w i0, i0, 1
ld.w t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
@@ -186,9 +299,46 @@
xvinsgr2vr.w VI0, i0, 6 //7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 7 //8
#endif
.align 3

.L24:
#ifdef DOUBLE
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
add.d X, X, INCX
ld.d t4, X, 0 * SIZE
add.d X, X, INCX
xvinsgr2vr.d VX0, t1, 0
xvinsgr2vr.d VX0, t2, 1
xvinsgr2vr.d VX0, t3, 2
xvinsgr2vr.d VX0, t4, 3
xvadd.d VI1, VI1, VINC8
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
add.d X, X, INCX
ld.d t4, X, 0 * SIZE
add.d X, X, INCX
xvinsgr2vr.d VX1, t1, 0
xvinsgr2vr.d VX1, t2, 1
xvinsgr2vr.d VX1, t3, 2
xvinsgr2vr.d VX1, t4, 3
xvadd.d VI2, VI1, VINC4
xvfmaxa.d VM1, VX0, VX1
xvfcmp.ceq.d VT0, VX0, VM1
addi.d I, I, -1
xvbitsel.v VI2, VI2, VI1, VT0
xvfmaxa.d VM1, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI2, VI0, VT0
#else
ld.w t1, X, 0 * SIZE
add.d X, X, INCX
ld.w t2, X, 0 * SIZE
@@ -219,10 +369,30 @@
addi.d I, I, -1
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
#endif
blt $r0, I, .L24
.align 3

.L25:
#ifdef DOUBLE
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
xvfmaxa.d VM1, x1, x2
xvfcmp.ceq.d VT0, x1, VM1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmaxa.d VM0, x4, x3
xvfcmp.ceq.d VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmaxa.d VM0, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
#else
xvxor.v VX0, VX0, VX0
xvor.v VX0, VI0, VX0
xvxor.v VX1, VX1, VX1
@@ -245,44 +415,45 @@
xvfcmp.ceq.s VT0, VM0, VM1
xvbitsel.v VM0, VM0, VM1, VT0
xvbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.w $f17, TEMP
ffint.s.w $f17, $f17
xvfcmp.ceq.s VT0, VM0, x1
fcmp.ceq.s $fcc0, $f23, $f17
#endif
CMPEQ $fcc0, $f15, $f9
bceqz $fcc0, .L26
xvfcmp.clt.s VT0, VI1, VI0
XVCMPLT VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
.align 3

.L26:
xvfcmp.ceq.s VT0, VM0, x2
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L27
xvfcmp.clt.s VT0, VI2, VI0
XVCMPLT VT0, VI2, VI0
xvbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
xvfcmp.ceq.s VT0, VM0, x3
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L28
xvfcmp.clt.s VT0, VI3, VI0
XVCMPLT VT0, VI3, VI0
xvbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
xvfcmp.ceq.s VT0, VM0, x4
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L29
xvfcmp.clt.s VT0, VI4, VI0
XVCMPLT VT0, VI4, VI0
xvbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
#ifdef DOUBLE
movfr2gr.d i0, $f20
#else
fmov.s $f16, $f20
#endif
.align 3

#ifdef DOUBLE

#else
.L252:
xvxor.v VI0, VI0, VI0
xvor.v VI0, VI0, VX0
@@ -306,35 +477,27 @@
xvfmaxa.s VM0, VM0, VM1
xvfcmp.ceq.s VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.w $f17, TEMP
ffint.s.w $f17, $f17
xvfcmp.ceq.s VT0, VM0, x1
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L262
xvfcmp.clt.s VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
.align 3

.L262:
xvfcmp.ceq.s VT0, VM0, x2
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L272
xvfcmp.clt.s VT0, VI2, VI0
xvbitsel.v VI0, VI0, VI2, VT0
.align 3

.L272:
xvfcmp.ceq.s VT0, VM0, x3
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L282
xvfcmp.clt.s VT0, VI3, VI0
xvbitsel.v VI0, VI0, VI3, VT0
.align 3

.L282:
xvfcmp.ceq.s VT0, VM0, x4
fcmp.ceq.s $fcc0, $f23, $f17
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L292
xvfcmp.clt.s VT0, VI4, VI0
xvbitsel.v VI0, VI0, VI4, VT0
@@ -345,8 +508,9 @@
xvfcmp.ceq.s VT0, VM0, VX0
xvbitsel.v VI0, VI0, VI1, VT0
movfr2gr.s i0, $f20
#endif

.L21: // N<8
.L21: //N<8
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
@@ -357,17 +521,17 @@
.align 3

.L22:
fld.s $f9, X, 0
LD $f9, X, 0
addi.d I, I, -1
xvfmaxa.s VM1, x1, VM0
xvfcmp.ceq.s VT0, VM0, VM1
XVFMAXA VM1, x1, VM0
XVCMPEQ VT0, VM0, VM1
add.d X, X, INCX
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
addi.d i1, i1, 1
movgr2fr.d $f21, i1
blt $r0, I, .L22
movfr2gr.s i0, $f20
MTG i0, $f20
.align 3

.L999:
@@ -375,4 +539,4 @@
jirl $r0, $r1, 0x0
.align 3

EPILOGUE
EPILOGUE

+ 482
- 0
kernel/loongarch64/iamax_lsx.S View File

@@ -0,0 +1,482 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define x1 $vr9
#define x2 $vr10
#define x3 $vr11
#define x4 $vr12
#define VX0 $vr13
#define VX1 $vr14
#define VM0 $vr15
#define VM1 $vr16
#define VINC2 $vr17
#define VINC4 $vr18
#define VI0 $vr20
#define VI1 $vr21
#define VI2 $vr22
#define VI3 $vr8
#define VI4 $vr19
#define VT0 $vr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT
bne INCX, TEMP, .L20
vld VM0, X, 0
#ifdef DOUBLE
addi.d i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L11
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC2, i0
slli.d i0, i0, 1 //4
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 3
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
#else
addi.w i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC2, i0
slli.w i0, i0, 1 //8
vreplgr2vr.w VINC4, i0
addi.w i0, i0, -15
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 5
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
#endif
.align 3

.L10:
#ifdef DOUBLE
vld VX0, X, 0 * SIZE
vadd.d VI1, VI1, VINC4
vld VX1, X, 2 * SIZE
vadd.d VI2, VI1, VINC2
vfmaxa.d x1, VX0, VX1
vfcmp.ceq.d VT0, VX0, x1
vbitsel.v x2, VI2, VI1, VT0
vld VX0, X, 4 * SIZE
vadd.d VI1, VI2, VINC2
vld VX1, X, 6 * SIZE
vadd.d VI2, VI1, VINC2
vfmaxa.d x3, VX0, VX1
vfcmp.ceq.d VT0, VX0, x3
vbitsel.v x4, VI2, VI1, VT0
vfmaxa.d x3, x1, x3
vfcmp.ceq.d VT0, x1, x3
vbitsel.v x2, x4, x2, VT0
vfmaxa.d VM1, VM0, x3
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, x2, VI0, VT0
addi.d I, I, -1
addi.d X, X, 8 * SIZE
#else
vld VX0, X, 0 * SIZE
vadd.w VI1, VI1, VINC4
vld VX1, X, 4 * SIZE
vadd.w VI2, VI1, VINC2
vfmaxa.s VM1, VX0, VX1
vfcmp.ceq.s VT0, VX0, VM1
addi.d I, I, -1
vbitsel.v VI2, VI2, VI1, VT0
vfmaxa.s VM1, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
addi.d X, X, 8 * SIZE
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI2, VI0, VT0
#endif
blt $r0, I, .L10
.align 3

.L15:
#ifdef DOUBLE
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
fcmp.ceq.d $fcc0, $f10, $f9
bceqz $fcc0, .L16
vfcmp.clt.d VT0, VI1, VI2
vbitsel.v VI0, VI2, VI1, VT0
b .L17
#else
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC2, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC4, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC4, VINC2, VT0
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
b .L26
#endif
.align 3

#ifdef DOUBLE
.L16:
vfmaxa.d VM0, x1, x2
vfcmp.ceq.d VT0, x1, VM0
vbitsel.v VI0, VI2, VI1, VT0
.align 3

.L17:
movfr2gr.d i0, $f20
.align 3

.L11: //INCX==1 and N<8
andi I, N, 7
bge $r0, I, .L14
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L13:
fld.d $f9, X, 0
vfmaxa.d VM1, x1, VM0
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
addi.d I, I, -1
addi.d i1, i1, 1
addi.d X, X, SIZE
movgr2fr.d $f21, i1
blt $r0, I, .L13
movfr2gr.d i0, $f20
.align 3

.L14:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

.L20: // INCX!=1
move TEMP, X
addi.d i0, i0, 1
ld.d t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.d VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.d t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.d VM0, t2, 1
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC2, i0
slli.d i0, i0, 1 //4
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 3
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
.align 3

.L24:
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t2, 1
vadd.d VI1, VI1, VINC4
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t2, 1
vadd.d VI2, VI1, VINC2
vfmaxa.d x1, VX0, VX1
vfcmp.ceq.d VT0, VX0, x1
vbitsel.v x2, VI2, VI1, VT0
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t2, 1
vadd.d VI1, VI2, VINC2
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t2, 1
vadd.d VI2, VI1, VINC2
vfmaxa.d x3, VX0, VX1
vfcmp.ceq.d VT0, VX0, x3
vbitsel.v x4, VI2, VI1, VT0
vfmaxa.d x3, x1, x3
vfcmp.ceq.d VT0, x1, x3
vbitsel.v x2, x4, x2, VT0
vfmaxa.d VM1, VM0, x3
vbitsel.v VM0, VM1, VM0, VT0
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VI0, x2, VI0, VT0
addi.d I, I, -1
blt $r0, I, .L24
.align 3

.L25:
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
fcmp.ceq.d $fcc0, $f10, $f9
bceqz $fcc0, .L26
vfcmp.clt.d VT0, VI1, VI2
vbitsel.v VI0, VI2, VI1, VT0
b .L27
.align 3

.L26:
vfmaxa.d VM0, x1, x2
vfcmp.ceq.d VT0, x1, VM0
vbitsel.v VI0, VI2, VI1, VT0
.align 3

.L27:
movfr2gr.d i0, $f20
.align 3

#else
.L20: // INCX!=1
move TEMP, X
addi.w i0, i0, 1
ld.w t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.w VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.w t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.w t3, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.w t4, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.w VM0, t2, 1
vinsgr2vr.w VM0, t3, 2
vinsgr2vr.w VM0, t4, 3
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC2, i0
slli.w i0, i0, 1 //8
vreplgr2vr.w VINC4, i0
addi.w i0, i0, -15
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 5
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
.align 3

.L24:
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 VX0, t1, 0
vinsgr2vr.w VX0, t2, 1
vinsgr2vr.w VX0, t3, 2
vinsgr2vr.w VX0, t4, 3
vadd.w VI1, VI1, VINC4
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
vadd.w VI2, VI1, VINC2
vfmaxa.s VM1, VX0, VX1
vfcmp.ceq.s VT0, VX0, VM1
vbitsel.v VI2, VI2, VI1, VT0
vfmaxa.s VM1, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
addi.d I, I, -1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI2, VI0, VT0
blt $r0, I, .L24
.align 3

.L25:
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC2, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC4, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC4, VINC2, VT0
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
.align 3

.L26:
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L27
vfcmp.clt.s VT0, VI2, VI0
vbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L28
vfcmp.clt.s VT0, VI3, VI0
vbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L29
vfcmp.clt.s VT0, VI4, VI0
vbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
movfr2gr.s i0, $f20
.align 3

#endif
.L21: // N<8
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
LD $f9, X, 0
VFMAXA VM1, x1, VM0
VCMPEQ VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
addi.d I, I, -1
addi.d i1, i1, 1
add.d X, X, INCX
movgr2fr.d $f21, i1
blt $r0, I, .L22
MTG i0, $f20
.align 3

.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 562
- 0
kernel/loongarch64/icamax_lasx.S View File

@@ -0,0 +1,562 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER
#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define a0 $f12
#define a1 $f13
#define s1 $f15
#define x1 $xr9
#define x2 $xr10
#define x3 $xr11
#define x4 $xr12
#define VX0 $xr13
#define VX1 $xr14
#define VM0 $xr15
#define VM1 $xr16
#define VINC4 $xr17
#define VINC8 $xr18
#define VI0 $xr20
#define VI1 $xr21
#define VI2 $xr22
#define VI3 $xr8
#define VI4 $xr19
#define VT0 $xr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
xvxor.v VM0, VM0, VM0
slli.d TEMP, TEMP, ZBASE_SHIFT
slli.d INCX, INCX, ZBASE_SHIFT
xvxor.v VI3, VI3, VI3 // 0
#ifdef DOUBLE
li.d I, -1
xvreplgr2vr.d VI4, I
xvffint.d.l VI4, VI4 // -1
bne INCX, TEMP, .L20
addi.d i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 2
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, -1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 2
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 2
xvinsgr2vr.d VI0, i0, 1 //3
addi.d i0, i0, -1
xvinsgr2vr.d VI0, i0, 2 //2
addi.d i0, i0, 2
xvinsgr2vr.d VI0, i0, 3 //4
#else
li.w I, -1
xvreplgr2vr.w VI4, I
xvffint.s.w VI4, VI4 // -1
bne INCX, TEMP, .L20
addi.w i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.w i0, i0, 3 //8
xvreplgr2vr.w VINC8, i0
addi.w i0, i0, -15
xvinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 3
xvinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 3
addi.w i0, i0, -3
xvinsgr2vr.w VI1, i0, 4
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 5
addi.w i0, i0, 3
xvinsgr2vr.w VI1, i0, 6
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 3
xvinsgr2vr.w VI0, i0, 2 //5
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 3 //6
addi.w i0, i0, -3
xvinsgr2vr.w VI0, i0, 4 //3
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 5 //4
addi.w i0, i0, 3
xvinsgr2vr.w VI0, i0, 6 //7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 7 //8
#endif
.align 3

.L10:
xvld VX0, X, 0 * SIZE
#ifdef DOUBLE
xvadd.d VI1, VI1, VINC4
xvld VX1, X, 4 * SIZE
addi.d I, I, -1
xvpickev.d x1, VX1, VX0
xvpickod.d x2, VX1, VX0
xvfmul.d x3, VI4, x1
xvfmul.d x4, VI4, x2
xvfcmp.clt.d VT0, x1, VI3
xvfcmp.clt.d VINC8, x2, VI3
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VINC8
#else
xvadd.w VI1, VI1, VINC8
xvld VX1, X, 8 * SIZE
addi.d I, I, -1
xvpickev.w x1, VX1, VX0
xvpickod.w x2, VX1, VX0
xvfmul.s x3, VI4, x1
xvfmul.s x4, VI4, x2
xvfcmp.clt.s VT0, x1, VI3
xvfcmp.clt.s VINC4, x2, VI3
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VINC4
#endif
XVFADD x1, x1, x2
XVFMAX x3, VM0, x1
XVCMPEQ VT0, x3, VM0
addi.d X, X, 8 * SIZE
xvbitsel.v VM0, x3, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
blt $r0, I, .L10
.align 3

.L15:
#ifdef DOUBLE
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
xvfmax.d VM1, x1, x2
xvfcmp.ceq.d VT0, VM1, x1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmax.d VM0, x3, x4
xvfcmp.ceq.d VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmax.d VM0, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
#else
xvxor.v VX0, VX0, VX0
xvor.v VX0, VI0, VX0
xvxor.v VX1, VX1, VX1
xvor.v VX1, VM0, VX1
xvpickve.w VI1, VI0, 0
xvpickve.w VI2, VI0, 1
xvpickve.w VI3, VI0, 2
xvpickve.w VI4, VI0, 3
xvpickve.w x1, VM0, 0
xvpickve.w x2, VM0, 1
xvpickve.w x3, VM0, 2
xvpickve.w x4, VM0, 3
xvfcmp.clt.s VT0, x1, x2
xvbitsel.v VM1, x1, x2, VT0
xvbitsel.v VINC4, VI1, VI2, VT0
xvfcmp.clt.s VT0, x3, x4
xvbitsel.v VM0, x3, x4, VT0
xvbitsel.v VINC8, VI3, VI4, VT0
xvfcmp.clt.s VT0, VM0, VM1
xvbitsel.v VM0, VM0, VM1, VT0
xvbitsel.v VI0, VINC8, VINC4, VT0
#endif
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
XVCMPLT VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
b .L26
.align 3

.L20: // INCX!=1
#ifdef DOUBLE
addi.d i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 2
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, -1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 2
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 2
xvinsgr2vr.d VI0, i0, 1 //3
addi.d i0, i0, -1
xvinsgr2vr.d VI0, i0, 2 //2
addi.d i0, i0, 2
xvinsgr2vr.d VI0, i0, 3 //4
#else
addi.w i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.w i0, i0, 3 //8
xvreplgr2vr.w VINC8, i0
addi.w i0, i0, -15
xvinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 3
xvinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 3
addi.w i0, i0, -3
xvinsgr2vr.w VI1, i0, 4
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 5
addi.w i0, i0, 3
xvinsgr2vr.w VI1, i0, 6
addi.w i0, i0, 1
xvinsgr2vr.w VI1, i0, 7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 3
xvinsgr2vr.w VI0, i0, 2 //5
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 3 //6
addi.w i0, i0, -3
xvinsgr2vr.w VI0, i0, 4 //3
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 5 //4
addi.w i0, i0, 3
xvinsgr2vr.w VI0, i0, 6 //7
addi.w i0, i0, 1
xvinsgr2vr.w VI0, i0, 7 //8
#endif
.align 3

.L24:
#ifdef DOUBLE
ld.d t1, X, 0 * SIZE
ld.d t2, X, 1 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
ld.d t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.d x1, t1, 0
xvinsgr2vr.d x2, t2, 0
xvinsgr2vr.d x1, t3, 1
xvinsgr2vr.d x2, t4, 1
xvadd.d VI1, VI1, VINC4
ld.d t1, X, 0 * SIZE
ld.d t2, X, 1 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
ld.d t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.d x1, t1, 2
xvinsgr2vr.d x2, t2, 2
xvinsgr2vr.d x1, t3, 3
xvinsgr2vr.d x2, t4, 3
addi.d I, I, -1
xvfmul.d x3, VI4, x1
xvfmul.d x4, VI4, x2
xvfcmp.clt.d VT0, x1, VI3
xvfcmp.clt.d VINC8, x2, VI3
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VINC8
xvfadd.d x1, x1, x2
xvfmax.d x3, VM0, x1
xvfcmp.ceq.d VT0, x3, VM0
#else
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.w x1, t1, 0
xvinsgr2vr.w x2, t2, 0
xvinsgr2vr.w x1, t3, 1
xvinsgr2vr.w x2, t4, 1
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.w x1, t1, 2
xvinsgr2vr.w x2, t2, 2
xvinsgr2vr.w x1, t3, 3
xvinsgr2vr.w x2, t4, 3
xvadd.w VI1, VI1, VINC8
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.w x1, t1, 4
xvinsgr2vr.w x2, t2, 4
xvinsgr2vr.w x1, t3, 5
xvinsgr2vr.w x2, t4, 5
xvadd.w VI1, VI1, VINC8
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
xvinsgr2vr.w x1, t1, 6
xvinsgr2vr.w x2, t2, 6
xvinsgr2vr.w x1, t3, 7
xvinsgr2vr.w x2, t4, 7
addi.d I, I, -1
xvpickev.w x1, VX1, VX0
xvpickod.w x2, VX1, VX0
xvfmul.s x3, VI4, x1
xvfmul.s x4, VI4, x2
xvfcmp.clt.s VT0, x1, VI3
xvfcmp.clt.s VINC8, x2, VI3
xvbitsel.v x1, x1, x3, VT0
xvbitsel.v x2, x2, x4, VINC8
xvfadd.s x1, x1, x2
xvfmax.s x3, VM0, x1
xvfcmp.ceq.s VT0, x3, VM0
#endif
xvbitsel.v VM0, x3, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
blt $r0, I, .L24
.align 3

.L25:
#ifdef DOUBLE
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
xvfmaxa.d VM1, x1, x2
xvfcmp.ceq.d VT0, VM1, x1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmaxa.d VM0, x3, x4
xvfcmp.ceq.d VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmaxa.d VM0, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
#else
xvxor.v VX0, VX0, VX0
xvor.v VX0, VI0, VX0
xvxor.v VX1, VX1, VX1
xvor.v VX1, VM0, VX1
xvpickve.w VI1, VI0, 0
xvpickve.w VI2, VI0, 1
xvpickve.w VI3, VI0, 2
xvpickve.w VI4, VI0, 3
xvpickve.w x1, VM0, 0
xvpickve.w x2, VM0, 1
xvpickve.w x3, VM0, 2
xvpickve.w x4, VM0, 3
xvfcmp.clt.s VT0, x1, x2
xvbitsel.v VM1, x1, x2, VT0
xvbitsel.v VINC4, VI1, VI2, VT0
xvfcmp.clt.s VT0, x3, x4
xvbitsel.v VM0, x3, x4, VT0
xvbitsel.v VINC8, VI3, VI4, VT0
xvfcmp.clt.s VT0, VM0, VM1
xvbitsel.v VM0, VM0, VM1, VT0
xvbitsel.v VI0, VINC8, VINC4, VT0
#endif
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
XVCMPLT VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
.align 3

.L26:
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L27
XVCMPLT VT0, VI2, VI0
xvbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L28
XVCMPLT VT0, VI3, VI0
xvbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L29
XVCMPLT VT0, VI4, VI0
xvbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
#ifdef DOUBLE
movfr2gr.d i0, $f20
#else
fmov.s $f16, $f20
#endif
.align 3

#ifdef DOUBLE
#else
.L252:
xvxor.v VI0, VI0, VI0
xvor.v VI0, VI0, VX0
fmov.s $f13, $f15
xvxor.v VM0, VM0, VM0
xvor.v VM0, VM0, VX1
xvpickve.w VI1, VI0, 4
xvpickve.w VI2, VI0, 5
xvpickve.w VI3, VI0, 6
xvpickve.w VI4, VI0, 7
xvpickve.w x1, VM0, 4
xvpickve.w x2, VM0, 5
xvpickve.w x3, VM0, 6
xvpickve.w x4, VM0, 7
xvfcmp.clt.s VT0, x1, x2
xvbitsel.v x1, x1, x2, VT0
xvbitsel.v VINC4, VI1, VI2, VT0
xvfcmp.clt.s VT0, x3, x4
xvbitsel.v VM0, x3, x4, VT0
xvbitsel.v VINC8, VI3, VI4, VT0
xvfcmp.clt.s VT0, VM0, x1
xvbitsel.v VM0, VM0, x1, VT0
xvbitsel.v VI0, VINC8, VINC4, VT0
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L262
xvfcmp.clt.s VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
.align 3

.L262:
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L272
xvfcmp.clt.s VT0, VI2, VI0
xvbitsel.v VI0, VI0, VI2, VT0
.align 3

.L272:
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L282
xvfcmp.clt.s VT0, VI3, VI0
xvbitsel.v VI0, VI0, VI3, VT0
.align 3

.L282:
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L292
xvfcmp.clt.s VT0, VI4, VI0
xvbitsel.v VI0, VI0, VI4, VT0
.align 3

.L292:
fcmp.clt.s $fcc0, $f15, $f13
fsel $f15, $f15, $f13, $fcc0
fsel $f20, $f20, $f16, $fcc0
movfr2gr.s i0, $f20

#endif
.L21: //N<8
#ifdef DOUBLE
andi I, N, 3
bge $r0, I, .L999
srai.d i1, N, 2
slli.d i1, i1, 2
#else
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
slli.d i1, i1, 3
#endif
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
LD a0, X, 0 * SIZE
LD a1, X, 1 * SIZE
addi.d I, I, -1
FABS a0, a0
FABS a1, a1
ADD a0, a0, a1
FMAX a1, s1, a0
CMPEQ $fcc0, s1, a1
add.d X, X, INCX
fsel s1, a1, s1, $fcc0
fsel $f20, $f21, $f20, $fcc0
addi.d i1, i1, 1
movgr2fr.d $f21, i1
blt $r0, I, .L22
MTG i0, $f20
.align 3


.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 434
- 0
kernel/loongarch64/icamax_lsx.S View File

@@ -0,0 +1,434 @@
/***************************************************************************
Copyright (c) 2023, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/

#define ASSEMBLER
#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define a0 $f12
#define a1 $f13
#define s1 $f15
#define x1 $vr9
#define x2 $vr10
#define x3 $vr11
#define x4 $vr12
#define VX0 $vr13
#define VX1 $vr14
#define VM0 $vr15
#define VM1 $vr16
#define VINC4 $vr17
#define VINC8 $vr18
#define VI0 $vr20
#define VI1 $vr21
#define VI2 $vr22
#define VI3 $vr8
#define VI4 $vr19
#define VT0 $vr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
vxor.v VM0, VM0, VM0
slli.d TEMP, TEMP, ZBASE_SHIFT
slli.d INCX, INCX, ZBASE_SHIFT
vxor.v VI3, VI3, VI3 // 0
#ifdef DOUBLE
li.d I, -1
vreplgr2vr.d VI4, I
vffint.d.l VI4, VI4 // -1
bne INCX, TEMP, .L20
addi.d i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -3
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
#else
li.w I, -1
vreplgr2vr.w VI4, I
vffint.s.w VI4, VI4 // -1
bne INCX, TEMP, .L20
addi.w i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC4, i0
addi.w i0, i0, -7
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
#endif
.align 3

.L10:
vld VX0, X, 0 * SIZE
#ifdef DOUBLE
vadd.d VI1, VI1, VINC4
vld VX1, X, 2 * SIZE
addi.d I, I, -1
vpickev.d x1, VX1, VX0
vpickod.d x2, VX1, VX0
vfmul.d x3, VI4, x1
vfmul.d x4, VI4, x2
vfcmp.clt.d VT0, x1, VI3
vfcmp.clt.d VINC8, x2, VI3
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VINC8
vfadd.d x1, x1, x2
vfmax.d x3, VM0, x1
vfcmp.ceq.d VT0, x3, VM0
vbitsel.v VM0, x3, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
vld VX0, X, 4 * SIZE
vadd.d VI1, VI1, VINC4
vld VX1, X, 6 * SIZE
vpickev.d x1, VX1, VX0
vpickod.d x2, VX1, VX0
vfmul.d x3, VI4, x1
vfmul.d x4, VI4, x2
#else
vadd.w VI1, VI1, VINC4
vld VX1, X, 4 * SIZE
addi.d I, I, -1
vpickev.w x1, VX1, VX0
vpickod.w x2, VX1, VX0
vfmul.s x3, VI4, x1
vfmul.s x4, VI4, x2
#endif
VCMPLT VT0, x1, VI3
VCMPLT VINC8, x2, VI3
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VINC8
VFADD x1, x1, x2
VFMAX x3, VM0, x1
VCMPEQ VT0, x3, VM0
addi.d X, X, 8 * SIZE
vbitsel.v VM0, x3, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
blt $r0, I, .L10
.align 3

.L15:
#ifdef DOUBLE
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
fcmp.ceq.d $fcc0, $f10, $f9
bceqz $fcc0, .L26
vfcmp.clt.d VT0, VI1, VI2
vbitsel.v VI0, VI2, VI1, VT0
b .L27
#else
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC4, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC8, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC8, VINC4, VT0
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
b .L26
#endif
.align 3

.L20: // INCX!=1
#ifdef DOUBLE
addi.d i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -3
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
#else
addi.w i0, i0, 1
srai.d I, N, 2
bge $r0, I, .L21
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC4, i0
addi.w i0, i0, -7
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
#endif
.align 3

.L24:
#ifdef DOUBLE
ld.d t1, X, 0 * SIZE
ld.d t2, X, 1 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
ld.d t4, X, 1 * SIZE
add.d X, X, INCX
vinsgr2vr.d x1, t1, 0
vinsgr2vr.d x2, t2, 0
vinsgr2vr.d x1, t3, 1
vinsgr2vr.d x2, t4, 1
vadd.d VI1, VI1, VINC4
vfmul.d x3, VI4, x1
vfmul.d x4, VI4, x2
vfcmp.clt.d VT0, x1, VI3
vfcmp.clt.d VINC8, x2, VI3
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VINC8
vfadd.d x1, x1, x2
vfmax.d x3, VM0, x1
ld.d t1, X, 0 * SIZE
vfcmp.ceq.d VT0, x3, VM0
ld.d t2, X, 1 * SIZE
vbitsel.v VM0, x3, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
ld.d t4, X, 1 * SIZE
add.d X, X, INCX
vinsgr2vr.d x1, t1, 0
vinsgr2vr.d x2, t2, 0
vinsgr2vr.d x1, t3, 1
vinsgr2vr.d x2, t4, 1
vadd.d VI1, VI1, VINC4
addi.d I, I, -1
vfmul.d x3, VI4, x1
vfmul.d x4, VI4, x2
vfcmp.clt.d VT0, x1, VI3
vfcmp.clt.d VINC8, x2, VI3
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VINC8
vfadd.d x1, x1, x2
vfmax.d x3, VM0, x1
vfcmp.ceq.d VT0, x3, VM0
#else
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
vinsgr2vr.w x1, t1, 0
vinsgr2vr.w x2, t2, 0
vinsgr2vr.w x1, t3, 1
vinsgr2vr.w x2, t4, 1
vadd.w VI1, VI1, VINC4
ld.w t1, X, 0 * SIZE
ld.w t2, X, 1 * SIZE
add.d X, X, INCX
ld.w t3, X, 0 * SIZE
ld.w t4, X, 1 * SIZE
add.d X, X, INCX
vinsgr2vr.w x1, t1, 2
vinsgr2vr.w x2, t2, 2
vinsgr2vr.w x1, t3, 3
vinsgr2vr.w x2, t4, 3
addi.d I, I, -1
vpickev.w x1, VX1, VX0
vpickod.w x2, VX1, VX0
vfmul.s x3, VI4, x1
vfmul.s x4, VI4, x2
vfcmp.clt.s VT0, x1, VI3
vfcmp.clt.s VINC8, x2, VI3
vbitsel.v x1, x1, x3, VT0
vbitsel.v x2, x2, x4, VINC8
vfadd.s x1, x1, x2
vfmax.s x3, VM0, x1
vfcmp.ceq.s VT0, x3, VM0
#endif
vbitsel.v VM0, x3, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
blt $r0, I, .L24
.align 3

.L25:
#ifdef DOUBLE
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
fcmp.ceq.d $fcc0, $f10, $f9
bceqz $fcc0, .L26
vfcmp.clt.d VT0, VI1, VI2
vbitsel.v VI0, VI2, VI1, VT0
b .L27
#else
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC4, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC8, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC8, VINC4, VT0
fcmp.ceq.d $fcc0, $f15, $f9
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
#endif
.align 3

#ifdef DOUBLE
.L26:
vfmaxa.d VM0, x1, x2
vfcmp.ceq.d VT0, x1, VM0
vbitsel.v VI0, VI2, VI1, VT0
.align 3

.L27:
movfr2gr.d i0, $f20
.align 3
#else
.L26:
fcmp.ceq.d $fcc0, $f15, $f10
bceqz $fcc0, .L27
vfcmp.clt.s VT0, VI2, VI0
vbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
fcmp.ceq.d $fcc0, $f15, $f11
bceqz $fcc0, .L28
vfcmp.clt.s VT0, VI3, VI0
vbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
fcmp.ceq.d $fcc0, $f15, $f12
bceqz $fcc0, .L29
vfcmp.clt.s VT0, VI4, VI0
vbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
movfr2gr.s i0, $f20
.align 3

#endif
.L21: //N<4
andi I, N, 3
bge $r0, I, .L999
srai.d i1, N, 2
slli.d i1, i1, 2
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
LD a0, X, 0 * SIZE
LD a1, X, 1 * SIZE
addi.d I, I, -1
FABS a0, a0
FABS a1, a1
ADD a0, a0, a1
FMAX a1, s1, a0
CMPEQ $fcc0, s1, a1
add.d X, X, INCX
fsel s1, a1, s1, $fcc0
fsel $f20, $f21, $f20, $fcc0
addi.d i1, i1, 1
movgr2fr.d $f21, i1
blt $r0, I, .L22
MTG i0, $f20
.align 3

.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 0
- 275
kernel/loongarch64/idamax_lasx.S View File

@@ -1,275 +0,0 @@
#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define x1 $xr9
#define x2 $xr10
#define x3 $xr11
#define x4 $xr12
#define VX0 $xr13
#define VX1 $xr14
#define VM0 $xr15
#define VM1 $xr16
#define VINC4 $xr17
#define VINC8 $xr18
#define VI0 $xr20
#define VI1 $xr21
#define VI2 $xr22
#define VI3 $xr8
#define VI4 $xr19
#define VT0 $xr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT
bne INCX, TEMP, .L20
xvld VM0, X, 0
addi.d i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
slli.d i0, i0, 1 //8
xvreplgr2vr.d VINC8, i0
addi.d i0, i0, -15
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 5
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 1 //2
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 2 //3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 3 //4
.align 3

.L10:
xvld VX0, X, 0 * SIZE
xvadd.d VI1, VI1, VINC8
xvld VX1, X, 4 * SIZE
xvadd.d VI2, VI1, VINC4
xvfmaxa.d VM1, VX0, VX1
xvfcmp.ceq.d VT0, VX0, VM1
addi.d I, I, -1
xvbitsel.v VI2, VI2, VI1, VT0
xvfmaxa.d VM1, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
addi.d X, X, 8 * SIZE
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI2, VI0, VT0
blt $r0, I, .L10
.align 3

.L15:
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
xvfmaxa.d VM1, x1, x2
xvfcmp.ceq.d VT0, x1, VM1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmaxa.d VM0, x4, x3
xvfcmp.ceq.d VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmaxa.d VM0, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.d $f17, TEMP
ffint.d.l $f17, $f17
xvfcmp.ceq.d VT0, VM0, x1
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L26
xvfcmp.clt.d VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
b .L26
.align 3

.L20: // INCX!=1
move TEMP, X
addi.d i0, i0, 1
ld.d t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
xvinsgr2vr.d VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.d t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.d t3, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.d t4, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
xvinsgr2vr.d VM0, t2, 1
xvinsgr2vr.d VM0, t3, 2
xvinsgr2vr.d VM0, t4, 3
slli.d i0, i0, 2 //4
xvreplgr2vr.d VINC4, i0
slli.d i0, i0, 1 //8
xvreplgr2vr.d VINC8, i0
addi.d i0, i0, -15
xvinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 2
addi.d i0, i0, 1
xvinsgr2vr.d VI1, i0, 3
addi.d i0, i0, 5
xvinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 1 //2
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 2 //3
addi.d i0, i0, 1
xvinsgr2vr.d VI0, i0, 3 //4
.align 3

.L24:
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
add.d X, X, INCX
ld.d t4, X, 0 * SIZE
add.d X, X, INCX
xvinsgr2vr.d VX0, t1, 0
xvinsgr2vr.d VX0, t2, 1
xvinsgr2vr.d VX0, t3, 2
xvinsgr2vr.d VX0, t4, 3
xvadd.d VI1, VI1, VINC8
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
ld.d t3, X, 0 * SIZE
add.d X, X, INCX
ld.d t4, X, 0 * SIZE
add.d X, X, INCX
xvinsgr2vr.d VX1, t1, 0
xvinsgr2vr.d VX1, t2, 1
xvinsgr2vr.d VX1, t3, 2
xvinsgr2vr.d VX1, t4, 3
xvadd.d VI2, VI1, VINC4
xvfmaxa.d VM1, VX0, VX1
xvfcmp.ceq.d VT0, VX0, VM1
addi.d I, I, -1
xvbitsel.v VI2, VI2, VI1, VT0
xvfmaxa.d VM1, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI2, VI0, VT0
blt $r0, I, .L24
.align 3

.L25:
xvpickve.d VI1, VI0, 0
xvpickve.d VI2, VI0, 1
xvpickve.d VI3, VI0, 2
xvpickve.d VI4, VI0, 3
xvpickve.d x1, VM0, 0
xvpickve.d x2, VM0, 1
xvpickve.d x3, VM0, 2
xvpickve.d x4, VM0, 3
xvfmaxa.d VM1, x1, x2
xvfcmp.ceq.d VT0, x1, VM1
xvbitsel.v VINC4, VI2, VI1, VT0
xvfmaxa.d VM0, x4, x3
xvfcmp.ceq.d VT0, x3, VM0
xvbitsel.v VINC8, VI4, VI3, VT0
xvfmaxa.d VM0, VM0, VM1
xvfcmp.ceq.d VT0, VM0, VM1
xvbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.d $f17, TEMP
ffint.d.l $f17, $f17
xvfcmp.ceq.d VT0, VM0, x1
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L26
xvfcmp.clt.d VT0, VI1, VI0
xvbitsel.v VI0, VI0, VI1, VT0
.align 3

.L26:
xvfcmp.ceq.d VT0, VM0, x2
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L27
xvfcmp.clt.d VT0, VI2, VI0
xvbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
xvfcmp.ceq.d VT0, VM0, x3
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L28
xvfcmp.clt.d VT0, VI3, VI0
xvbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
xvfcmp.ceq.d VT0, VM0, x4
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L29
xvfcmp.clt.d VT0, VI4, VI0
xvbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
movfr2gr.d i0, $f20
.align 3

.L21: //N<8
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
fld.d $f9, X, 0
addi.d I, I, -1
xvfmaxa.d VM1, x1, VM0
xvfcmp.ceq.d VT0, VM0, VM1
add.d X, X, INCX
xvbitsel.v VM0, VM1, VM0, VT0
xvbitsel.v VI0, VI1, VI0, VT0
addi.d i1, i1, 1
movgr2fr.d $f21, i1
blt $r0, I, .L22
movfr2gr.d i0, $f20
.align 3

.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 0
- 267
kernel/loongarch64/idamax_lsx.S View File

@@ -1,267 +0,0 @@
#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define x1 $vr9
#define x2 $vr10
#define x3 $vr11
#define x4 $vr12
#define VX0 $vr13
#define VX1 $vr14
#define VM0 $vr15
#define VM1 $vr16
#define VINC2 $vr17
#define VINC4 $vr18
#define VI0 $vr20
#define VI1 $vr21
#define VI2 $vr22
#define VI3 $vr8
#define VI4 $vr19
#define VT0 $vr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT
bne INCX, TEMP, .L20
vld VM0, X, 0
addi.d i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L11
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC2, i0
slli.d i0, i0, 1 //4
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 3
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
.align 3

.L10:
vld VX0, X, 0 * SIZE
vadd.d VI1, VI1, VINC4
vld VX1, X, 2 * SIZE
vadd.d VI2, VI1, VINC2
vfmaxa.d x1, VX0, VX1
vfcmp.ceq.d VT0, VX0, x1
vbitsel.v x2, VI2, VI1, VT0
vld VX0, X, 4 * SIZE
vadd.d VI1, VI2, VINC2
vld VX1, X, 6 * SIZE
vadd.d VI2, VI1, VINC2
vfmaxa.d x3, VX0, VX1
vfcmp.ceq.d VT0, VX0, x3
vbitsel.v x4, VI2, VI1, VT0
vfmaxa.d x3, x1, x3
vfcmp.ceq.d VT0, x1, x3
vbitsel.v x2, x4, x2, VT0
vfmaxa.d VM1, VM0, x3
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, x2, VI0, VT0
addi.d I, I, -1
addi.d X, X, 8 * SIZE
blt $r0, I, .L10
.align 3

.L15:
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.d $f17, TEMP
ffint.d.l $f17, $f17
vfcmp.ceq.d VT0, x2, x1
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L16
vfcmp.clt.d VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
b .L17
.align 3

.L16:
vfmaxa.d VM0, x1, x2
vfcmp.ceq.d VT0, x1, VM0
vbitsel.v VI0, VI2, VI1, VT0
.align 3

.L17:
movfr2gr.d i0, $f20
.align 3

.L11: //INCX==1 and N<8
andi I, N, 7
bge $r0, I, .L14
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L13:
fld.d $f9, X, 0
vfmaxa.d VM1, x1, VM0
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
addi.d I, I, -1
addi.d i1, i1, 1
addi.d X, X, SIZE
movgr2fr.d $f21, i1
blt $r0, I, .L13
movfr2gr.d i0, $f20
.align 3

.L14:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

.L20: // INCX!=1
move TEMP, X
addi.d i0, i0, 1
ld.d t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.d VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.d t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.d VM0, t2, 1
slli.d i0, i0, 1 //2
vreplgr2vr.d VINC2, i0
slli.d i0, i0, 1 //4
vreplgr2vr.d VINC4, i0
addi.d i0, i0, -7
vinsgr2vr.d VI1, i0, 0 //initialize the index value for vectorization
addi.d i0, i0, 1
vinsgr2vr.d VI1, i0, 1
addi.d i0, i0, 3
vinsgr2vr.d VI0, i0, 0 //1
addi.d i0, i0, 1
vinsgr2vr.d VI0, i0, 1 //2
.align 3

.L24:
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t2, 1
vadd.d VI1, VI1, VINC4
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t2, 1
vadd.d VI2, VI1, VINC2
vfmaxa.d x1, VX0, VX1
vfcmp.ceq.d VT0, VX0, x1
vbitsel.v x2, VI2, VI1, VT0
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX0, t2, 1
vadd.d VI1, VI2, VINC2
ld.d t1, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t1, 0
ld.d t2, X, 0 * SIZE
add.d X, X, INCX
vinsgr2vr.d VX1, t2, 1
vadd.d VI2, VI1, VINC2
vfmaxa.d x3, VX0, VX1
vfcmp.ceq.d VT0, VX0, x3
vbitsel.v x4, VI2, VI1, VT0
vfmaxa.d x3, x1, x3
vfcmp.ceq.d VT0, x1, x3
vbitsel.v x2, x4, x2, VT0
vfmaxa.d VM1, VM0, x3
vbitsel.v VM0, VM1, VM0, VT0
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VI0, x2, VI0, VT0
addi.d I, I, -1
blt $r0, I, .L24
.align 3

.L25:
vreplvei.d VI1, VI0, 0
vreplvei.d VI2, VI0, 1
vreplvei.d x1, VM0, 0
vreplvei.d x2, VM0, 1
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.d $f17, TEMP
ffint.d.l $f17, $f17
vfcmp.ceq.d VT0, x2, x1
fcmp.ceq.d $fcc0, $f23, $f17
bceqz $fcc0, .L26
vfcmp.clt.d VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
b .L27
.align 3

.L26:
vfmaxa.d VM0, x1, x2
vfcmp.ceq.d VT0, x1, VM0
vbitsel.v VI0, VI2, VI1, VT0
.align 3

.L27:
movfr2gr.d i0, $f20
.align 3

.L21: // N<8
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
fld.d $f9, X, 0
vfmaxa.d VM1, x1, VM0
vfcmp.ceq.d VT0, VM0, VM1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
addi.d I, I, -1
addi.d i1, i1, 1
add.d X, X, INCX
movgr2fr.d $f21, i1
blt $r0, I, .L22
movfr2gr.d i0, $f20
.align 3

.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

+ 0
- 275
kernel/loongarch64/isamax_lsx.S View File

@@ -1,275 +0,0 @@
#define ASSEMBLER

#include "common.h"

#define N $r4
#define X $r5
#define INCX $r6
#define I $r12
#define t1 $r13
#define t2 $r15
#define t3 $r18
#define t4 $r16
#define i0 $r17
#define i1 $r14
#define TEMP $r19
#define x1 $vr9
#define x2 $vr10
#define x3 $vr11
#define x4 $vr12
#define VX0 $vr13
#define VX1 $vr14
#define VM0 $vr15
#define VM1 $vr16
#define VINC4 $vr17
#define VINC8 $vr18
#define VI0 $vr20
#define VI1 $vr21
#define VI2 $vr22
#define VI3 $vr8
#define VI4 $vr19
#define VT0 $vr23

PROLOGUE
li.d i0, 0
bge $r0, N, .L999
bge $r0, INCX, .L999
li.d TEMP, 1
slli.d TEMP, TEMP, BASE_SHIFT
slli.d INCX, INCX, BASE_SHIFT
bne INCX, TEMP, .L20
vld VM0, X, 0
addi.w i0, i0, 1
srai.d I, N, 3
bge $r0, I, .L21
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC4, i0
slli.w i0, i0, 1 //8
vreplgr2vr.w VINC8, i0
addi.w i0, i0, -15
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 5
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
.align 3

.L10:
vld VX0, X, 0 * SIZE
vadd.w VI1, VI1, VINC8
vld VX1, X, 4 * SIZE
vadd.w VI2, VI1, VINC4
vfmaxa.s VM1, VX0, VX1
vfcmp.ceq.s VT0, VX0, VM1
addi.d I, I, -1
vbitsel.v VI2, VI2, VI1, VT0
vfmaxa.s VM1, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
addi.d X, X, 8 * SIZE
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI2, VI0, VT0
blt $r0, I, .L10
.align 3

.L15:
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC4, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC8, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.w $f17, TEMP
ffint.s.w $f17, $f17
vfcmp.ceq.s VT0, VM0, x1
fcmp.ceq.s $fcc0, $f23, $f17
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
b .L26
.align 3

.L20: // INCX!=1
move TEMP, X
addi.w i0, i0, 1
ld.w t1, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.w VM0, t1, 0
srai.d I, N, 3
bge $r0, I, .L21
ld.w t2, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.w t3, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
ld.w t4, TEMP, 0 * SIZE
add.d TEMP, TEMP, INCX
vinsgr2vr.w VM0, t2, 1
vinsgr2vr.w VM0, t3, 2
vinsgr2vr.w VM0, t4, 3
slli.w i0, i0, 2 //4
vreplgr2vr.w VINC4, i0
slli.w i0, i0, 1 //8
vreplgr2vr.w VINC8, i0
addi.w i0, i0, -15
vinsgr2vr.w VI1, i0, 0 //initialize the index value for vectorization
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 1
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 2
addi.w i0, i0, 1
vinsgr2vr.w VI1, i0, 3
addi.w i0, i0, 5
vinsgr2vr.w VI0, i0, 0 //1
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 1 //2
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 2 //3
addi.w i0, i0, 1
vinsgr2vr.w VI0, i0, 3 //4
.align 3

.L24:
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 VX0, t1, 0
vinsgr2vr.w VX0, t2, 1
vinsgr2vr.w VX0, t3, 2
vinsgr2vr.w VX0, t4, 3
vadd.w VI1, VI1, VINC8
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
vadd.w VI2, VI1, VINC4
vfmaxa.s VM1, VX0, VX1
vfcmp.ceq.s VT0, VX0, VM1
vbitsel.v VI2, VI2, VI1, VT0
vfmaxa.s VM1, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
addi.d I, I, -1
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI2, VI0, VT0
blt $r0, I, .L24
.align 3

.L25:
vreplvei.w VI1, VI0, 0
vreplvei.w VI2, VI0, 1
vreplvei.w VI3, VI0, 2
vreplvei.w VI4, VI0, 3
vreplvei.w x1, VM0, 0
vreplvei.w x2, VM0, 1
vreplvei.w x3, VM0, 2
vreplvei.w x4, VM0, 3
vfmaxa.s VM1, x1, x2
vfcmp.ceq.s VT0, VM1, x1
vbitsel.v VINC4, VI2, VI1, VT0
vfmaxa.s VM0, x3, x4
vfcmp.ceq.s VT0, x3, VM0
vbitsel.v VINC8, VI4, VI3, VT0
vfmaxa.s VM0, VM0, VM1
vfcmp.ceq.s VT0, VM0, VM1
vbitsel.v VI0, VINC8, VINC4, VT0
li.d TEMP, 1 //处理尾数相等时取最小序号
movgr2fr.w $f17, TEMP
ffint.s.w $f17, $f17
vfcmp.ceq.s VT0, VM0, x1
fcmp.ceq.s $fcc0, $f23, $f17
bceqz $fcc0, .L26
vfcmp.clt.s VT0, VI1, VI0
vbitsel.v VI0, VI0, VI1, VT0
.align 3

.L26:
vfcmp.ceq.s VT0, VM0, x2
fcmp.ceq.s $fcc0, $f23, $f17
bceqz $fcc0, .L27
vfcmp.clt.s VT0, VI2, VI0
vbitsel.v VI0, VI0, VI2, VT0
.align 3

.L27:
vfcmp.ceq.s VT0, VM0, x3
fcmp.ceq.s $fcc0, $f23, $f17
bceqz $fcc0, .L28
vfcmp.clt.s VT0, VI3, VI0
vbitsel.v VI0, VI0, VI3, VT0
.align 3

.L28:
vfcmp.ceq.s VT0, VM0, x4
fcmp.ceq.s $fcc0, $f23, $f17
bceqz $fcc0, .L29
vfcmp.clt.s VT0, VI4, VI0
vbitsel.v VI0, VI0, VI4, VT0
.align 3

.L29:
movfr2gr.s i0, $f20
.align 3

.L21: //N<8
andi I, N, 7
bge $r0, I, .L999
srai.d i1, N, 3
slli.d i1, i1, 3
addi.d i1, i1, 1 //current index
movgr2fr.d $f21, i1
movgr2fr.d $f20, i0
.align 3

.L22:
fld.s $f9, X, 0
addi.d I, I, -1
vfmaxa.s VM1, x1, VM0
vfcmp.ceq.s VT0, VM0, VM1
add.d X, X, INCX
vbitsel.v VM0, VM1, VM0, VT0
vbitsel.v VI0, VI1, VI0, VT0
addi.d i1, i1, 1
movgr2fr.d $f21, i1
blt $r0, I, .L22
movfr2gr.s i0, $f20
.align 3

.L999:
move $r4, $r17
jirl $r0, $r1, 0x0
.align 3

EPILOGUE

Write Preview
Loading…
Cancel
Save