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: Fixed dot_lsx.S 

Fixed incorrect register usage in instructions

Signed-off-by: gxw <guxiwei-hf@loongson.cn>
tags/v0.3.30
Hao Chen gxw 1 year ago
parent
5d6356bc16
commit
31d326f895
1 changed files with 29 additions and 55 deletions
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +29
    -55
      kernel/loongarch64/dot_lsx.S

+ 29
- 55
kernel/loongarch64/dot_lsx.S View File

@@ -53,8 +53,8 @@ PROLOGUE
#endif #endif


/* init $f8 and $f9 to zero */ /* init $f8 and $f9 to zero */
SUB s1, s1, s1
SUB s2, s2, s2
vxor.v $vr8, $vr8, $vr8
vxor.v $vr9, $vr9, $vr9
slli.d INCX, INCX, BASE_SHIFT slli.d INCX, INCX, BASE_SHIFT
li.d TEMP, SIZE li.d TEMP, SIZE
slli.d INCY, INCY, BASE_SHIFT slli.d INCY, INCY, BASE_SHIFT
@@ -64,20 +64,6 @@ PROLOGUE


/* !((inc_x == 1) && (inc_y == 1)) */ /* !((inc_x == 1) && (inc_y == 1)) */


/* init $vr8 and $vr9 to zero */
#ifdef DOUBLE
vldrepl.d $vr0, X, 0
#else
vldrepl.w $vr0, X, 0
#endif
#ifdef DSDOT
vfcvtl.d.s $vr0, $vr0
vfsub.d $vr8, $vr0, $vr0
vfsub.d $vr9, $vr0, $vr0
#else
VFSUB $vr8, $vr0, $vr0
VFSUB $vr9, $vr0, $vr0
#endif


#ifdef DOUBLE #ifdef DOUBLE
srai.d I, N, 3 srai.d I, N, 3
@@ -99,31 +85,31 @@ PROLOGUE
addi.w I, I, -1 addi.w I, I, -1
addi.d X, X, 64 addi.d X, X, 64
addi.d Y, Y, 64 addi.d Y, Y, 64
#ifdef DSDOT
#ifndef DOUBLE
vfcvtl.d.s $vr10, $vr0 vfcvtl.d.s $vr10, $vr0
vfcvtl.d.s $vr11, $vr4 vfcvtl.d.s $vr11, $vr4
vfcvth.d.s $vr12, $vr0 vfcvth.d.s $vr12, $vr0
vfcvth.d.s $vr13, $vr4 vfcvth.d.s $vr13, $vr4
vfmadd.d $vr8, $vr10, $vr12, $vr8
vfmadd.d $vr9, $vr11, $vr13, $vr9
vfmadd.d $vr8, $vr10, $vr11, $vr8
vfmadd.d $vr9, $vr12, $vr13, $vr9
vfcvtl.d.s $vr10, $vr1 vfcvtl.d.s $vr10, $vr1
vfcvtl.d.s $vr11, $vr5 vfcvtl.d.s $vr11, $vr5
vfcvth.d.s $vr12, $vr1 vfcvth.d.s $vr12, $vr1
vfcvth.d.s $vr13, $vr5 vfcvth.d.s $vr13, $vr5
vfmadd.d $vr8, $vr10, $vr12, $vr8
vfmadd.d $vr9, $vr11, $vr13, $vr9
vfmadd.d $vr8, $vr10, $vr11, $vr8
vfmadd.d $vr9, $vr12, $vr13, $vr9
vfcvtl.d.s $vr10, $vr2 vfcvtl.d.s $vr10, $vr2
vfcvtl.d.s $vr11, $vr6 vfcvtl.d.s $vr11, $vr6
vfcvth.d.s $vr12, $vr2 vfcvth.d.s $vr12, $vr2
vfcvth.d.s $vr13, $vr6 vfcvth.d.s $vr13, $vr6
vfmadd.d $vr8, $vr10, $vr12, $vr8
vfmadd.d $vr9, $vr11, $vr13, $vr9
vfmadd.d $vr8, $vr10, $vr11, $vr8
vfmadd.d $vr9, $vr12, $vr13, $vr9
vfcvtl.d.s $vr10, $vr3 vfcvtl.d.s $vr10, $vr3
vfcvtl.d.s $vr11, $vr7 vfcvtl.d.s $vr11, $vr7
vfcvth.d.s $vr12, $vr3 vfcvth.d.s $vr12, $vr3
vfcvth.d.s $vr13, $vr7 vfcvth.d.s $vr13, $vr7
vfmadd.d $vr8, $vr10, $vr12, $vr8
vfmadd.d $vr9, $vr11, $vr13, $vr9
vfmadd.d $vr8, $vr10, $vr11, $vr8
vfmadd.d $vr9, $vr12, $vr13, $vr9
#else #else
VFMADD $vr8, $vr0, $vr4, $vr8 VFMADD $vr8, $vr0, $vr4, $vr8
VFMADD $vr9, $vr1, $vr5, $vr9 VFMADD $vr9, $vr1, $vr5, $vr9
@@ -149,13 +135,13 @@ PROLOGUE
addi.w I, I, -1 addi.w I, I, -1
addi.d X, X, 16 addi.d X, X, 16
addi.d Y, Y, 16 addi.d Y, Y, 16
#ifdef DSDOT
#ifndef DOUBLE
vfcvtl.d.s $vr10, $vr0 vfcvtl.d.s $vr10, $vr0
vfcvtl.d.s $vr11, $vr4 vfcvtl.d.s $vr11, $vr4
vfcvth.d.s $vr12, $vr0 vfcvth.d.s $vr12, $vr0
vfcvth.d.s $vr13, $vr4 vfcvth.d.s $vr13, $vr4
vfmadd.d $vr8, $vr10, $vr12, $vr8
vfmadd.d $vr9, $vr11, $vr13, $vr9
vfmadd.d $vr8, $vr10, $vr11, $vr8
vfmadd.d $vr9, $vr12, $vr13, $vr9
#else #else
VFMADD $vr8, $vr0, $vr4, $vr8 VFMADD $vr8, $vr0, $vr4, $vr8
#endif #endif
@@ -163,23 +149,10 @@ PROLOGUE
.align 3 .align 3
.L14: .L14:
/* store dot in s1 $f8 */ /* store dot in s1 $f8 */
#ifdef DSDOT
vfadd.d $vr8, $vr8, $vr9 vfadd.d $vr8, $vr8, $vr9
fsub.s s2, s2, s2 /* set s2 to 0.0 */
fsub.d s2, s2, s2 /* set s2 to 0.0 */
vpackod.d $vr0, $vr8, $vr8 vpackod.d $vr0, $vr8, $vr8
vfadd.d $vr8, $vr8, $vr0 vfadd.d $vr8, $vr8, $vr0
#else
VFADD $vr8, $vr8, $vr9
SUB s2, s2, s2 /* set s2 to 0.0 */
vpackod.d $vr0, $vr8, $vr8
#ifdef DOUBLE
VFADD $vr8, $vr8, $vr0
#else
VFADD $vr8, $vr8, $vr0
vpackod.w $vr0, $vr8, $vr8
VFADD $vr8, $vr8, $vr0
#endif /* defined DOUBLE */
#endif /* defined DSDOT */
.align 3 .align 3
.L15: .L15:
#ifdef DOUBLE #ifdef DOUBLE
@@ -193,7 +166,7 @@ PROLOGUE
/* DOUBLE: 1 ; FLOAT: 1~3 */ /* DOUBLE: 1 ; FLOAT: 1~3 */
LD a1, X, 0 LD a1, X, 0
LD b1, Y, 0 LD b1, Y, 0
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -236,7 +209,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -248,7 +221,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s2, b1, a1, s2 fmadd.d s2, b1, a1, s2
@@ -260,7 +233,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -272,7 +245,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s2, b1, a1, s2 fmadd.d s2, b1, a1, s2
@@ -284,7 +257,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -296,7 +269,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s2, b1, a1, s2 fmadd.d s2, b1, a1, s2
@@ -308,7 +281,7 @@ PROLOGUE
add.d X, X, INCX add.d X, X, INCX
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -321,7 +294,7 @@ PROLOGUE
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
addi.d I, I, -1 addi.d I, I, -1
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s2, b1, a1, s2 fmadd.d s2, b1, a1, s2
@@ -342,7 +315,7 @@ PROLOGUE
LD b1, Y, 0 * SIZE LD b1, Y, 0 * SIZE
add.d Y, Y, INCY add.d Y, Y, INCY
addi.d I, I, -1 addi.d I, I, -1
#ifdef DSDOT
#ifndef DOUBLE
fcvt.d.s a1, a1 fcvt.d.s a1, a1
fcvt.d.s b1, b1 fcvt.d.s b1, b1
fmadd.d s1, b1, a1, s1 fmadd.d s1, b1, a1, s1
@@ -353,12 +326,13 @@ PROLOGUE
.align 3 .align 3


.L999: .L999:
#ifdef DSDOT
fadd.d $f0, s1, s2 fadd.d $f0, s1, s2
move $r4, $r17
#if defined(DOUBLE)
#elif defined(DSDOT)
#else #else
ADD $f0, s1, s2
fcvt.s.d $f0, $f0
#endif #endif
move $r4, $r17
jirl $r0, $r1, 0x0 jirl $r0, $r1, 0x0


EPILOGUE EPILOGUE

Write Preview
Loading…
Cancel
Save