floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #2312 from martin-frbg/power8be 

Further Power8 big-endian corrections
This commit is contained in:
Martin Kroeker 2019-11-20 15:12:06 +01:00 committed by GitHub
parent 63d3ee8dfc cad0d150db
commit 08fa83aba2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
8 changed files with 42 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
kernel/power/KERNEL.POWER8
View File

@ -89,14 +89,30 @@ ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
#SMINKERNEL = ../arm/min.c #SMINKERNEL = ../arm/min.c
#DMINKERNEL = ../arm/min.c #DMINKERNEL = ../arm/min.c
# #
ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
ISAMAXKERNEL = isamax_power8.S ISAMAXKERNEL = isamax_power8.S
else
ISAMAXKERNEL = isamax.c
endif
IDAMAXKERNEL = idamax.c IDAMAXKERNEL = idamax.c
ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
ICAMAXKERNEL = icamax_power8.S ICAMAXKERNEL = icamax_power8.S
else
ICAMAXKERNEL = icamax.c
endif
IZAMAXKERNEL = izamax.c IZAMAXKERNEL = izamax.c
# #
ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
ISAMINKERNEL = isamin_power8.S ISAMINKERNEL = isamin_power8.S
else
ISAMINKERNEL = isamin.c
endif
IDAMINKERNEL = idamin.c IDAMINKERNEL = idamin.c
ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
ICAMINKERNEL = icamin_power8.S ICAMINKERNEL = icamin_power8.S
else
ICAMINKERNEL = icamin.c
endif
IZAMINKERNEL = izamin.c IZAMINKERNEL = izamin.c
# #
#ISMAXKERNEL = ../arm/imax.c #ISMAXKERNEL = ../arm/imax.c
@ -112,7 +128,11 @@ ZASUMKERNEL = zasum.c
# #
SAXPYKERNEL = saxpy.c SAXPYKERNEL = saxpy.c
DAXPYKERNEL = daxpy.c DAXPYKERNEL = daxpy.c
ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
CAXPYKERNEL = caxpy_power8.S CAXPYKERNEL = caxpy_power8.S
else
CAXPYKERNEL = caxpy.c
endif
ZAXPYKERNEL = zaxpy.c ZAXPYKERNEL = zaxpy.c
# #
SCOPYKERNEL = scopy.c SCOPYKERNEL = scopy.c

5
kernel/power/caxpy_power8.S
View File

@ -12,11 +12,12 @@
PROLOGUE PROLOGUE
caxpy_k:
.LCF0: .LCF0:
0: addis 2,12,.TOC.-.LCF0@ha 0: addis 2,12,.TOC.-.LCF0@ha
addi 2,2,.TOC.-.LCF0@l addi 2,2,.TOC.-.LCF0@l
#if _CALL_ELF ==2
.localentry caxpy_k,.-caxpy_k .localentry caxpy_k,.-caxpy_k
#endif
mr. 7,3 mr. 7,3
ble 0,.L33 ble 0,.L33
cmpdi 7,9,1 cmpdi 7,9,1
@ -515,7 +516,9 @@ caxpy_k:
b .L13 b .L13
.long 0 .long 0
.byte 0,0,0,0,0,4,0,0 .byte 0,0,0,0,0,4,0,0
#if _CALL_ELF ==2
.size caxpy_k,.-caxpy_k .size caxpy_k,.-caxpy_k
#endif
.section .rodata .section .rodata
.align 4 .align 4
.set .LANCHOR0,. + 0 .set .LANCHOR0,. + 0

5
kernel/power/icamin_power8.S
View File

@ -11,11 +11,12 @@
PROLOGUE PROLOGUE
icamin_k:
.LCF0: .LCF0:
0: addis 2,12,.TOC.-.LCF0@ha 0: addis 2,12,.TOC.-.LCF0@ha
addi 2,2,.TOC.-.LCF0@l addi 2,2,.TOC.-.LCF0@l
#if _CALL_ELF ==2
.localentry icamin_k,.-icamin_k .localentry icamin_k,.-icamin_k
#endif
mr. 9,3 mr. 9,3
ble 0,.L25 ble 0,.L25
cmpdi 7,5,0 cmpdi 7,5,0
@ -388,7 +389,9 @@ icamin_k:
b .L21 b .L21
.long 0 .long 0
.byte 0,0,0,0,0,1,0,0 .byte 0,0,0,0,0,1,0,0
#if _CALL_ELF ==2
.size icamin_k,.-icamin_k .size icamin_k,.-icamin_k
#endif
.section .rodata.cst16,"aM",@progbits,16 .section .rodata.cst16,"aM",@progbits,16
.align 4 .align 4
.LC2: .LC2:

6
kernel/power/idamax.c
View File

@ -324,15 +324,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
if (inc_x == 1) { if (inc_x == 1) {
#if defined(_CALL_ELF) && (_CALL_ELF == 2)
BLASLONG n1 = n & -32; BLASLONG n1 = n & -32;
if (n1 > 0) { #if defined(_CALL_ELF) && (_CALL_ELF == 2)
if (n1 > 0) {
max = diamax_kernel_32(n1, x, &maxf); max = diamax_kernel_32(n1, x, &maxf);
i = n1; i = n1;
} }
#endif #endif
while (i < n) { while (i < n) {
if (ABS(x[i]) > maxf) { if (ABS(x[i]) > maxf) {
max = i; max = i;

3
kernel/power/idamin.c
View File

@ -328,13 +328,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
#if defined(_CALL_ELF) && (_CALL_ELF == 2) #if defined(_CALL_ELF) && (_CALL_ELF == 2)
BLASLONG n1 = n & -32; BLASLONG n1 = n & -32;
if (n1 > 0) { if (n1 > 0) {
min = diamin_kernel_32(n1, x, &minf); min = diamin_kernel_32(n1, x, &minf);
i = n1; i = n1;
} }
#endif #endif
while (i < n) { while (i < n) {
if (ABS(x[i]) < minf) { if (ABS(x[i]) < minf) {
min = i; min = i;

5
kernel/power/isamax_power8.S
View File

@ -12,11 +12,12 @@
PROLOGUE PROLOGUE
isamax_k:
.LCF0: .LCF0:
0: addis 2,12,.TOC.-.LCF0@ha 0: addis 2,12,.TOC.-.LCF0@ha
addi 2,2,.TOC.-.LCF0@l addi 2,2,.TOC.-.LCF0@l
#if _CALL_ELF ==2
.localentry isamax_k,.-isamax_k .localentry isamax_k,.-isamax_k
#endif
mr. 11,3 mr. 11,3
ble 0,.L36 ble 0,.L36
cmpdi 7,5,0 cmpdi 7,5,0
@ -397,7 +398,9 @@ isamax_k:
b .L61 b .L61
.long 0 .long 0
.byte 0,0,0,0,0,1,0,0 .byte 0,0,0,0,0,1,0,0
#if _CALL_ELF ==2
.size isamax_k,.-isamax_k .size isamax_k,.-isamax_k
#endif
.section .rodata.cst16,"aM",@progbits,16 .section .rodata.cst16,"aM",@progbits,16
.align 4 .align 4
.LC2: .LC2:

5
kernel/power/isamin_power8.S
View File

@ -11,11 +11,12 @@
PROLOGUE PROLOGUE
isamin_k:
.LCF0: .LCF0:
0: addis 2,12,.TOC.-.LCF0@ha 0: addis 2,12,.TOC.-.LCF0@ha
addi 2,2,.TOC.-.LCF0@l addi 2,2,.TOC.-.LCF0@l
#if _CALL_ELF ==2
.localentry isamin_k,.-isamin_k .localentry isamin_k,.-isamin_k
#endif
mr. 11,3 mr. 11,3
ble 0,.L36 ble 0,.L36
cmpdi 7,5,0 cmpdi 7,5,0
@ -380,7 +381,9 @@ isamin_k:
b .L35 b .L35
.long 0 .long 0
.byte 0,0,0,0,0,1,0,0 .byte 0,0,0,0,0,1,0,0
#if _CALL_ELF ==2
.size isamin_k,.-isamin_k .size isamin_k,.-isamin_k
#endif
.section .rodata.cst16,"aM",@progbits,16 .section .rodata.cst16,"aM",@progbits,16
.align 4 .align 4
.LC2: .LC2:

4
kernel/power/izamin.c
View File

@ -316,14 +316,14 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
minf = CABS1(x,0); //index will not be incremented minf = CABS1(x,0); //index will not be incremented
#if defined(_CALL_ELF) && (_CALL_ELF == 2) #if defined(_CALL_ELF) && (_CALL_ELF == 2)
BLASLONG n1 = n & -16; BLASLONG n1 = n & -16;
if (n1 > 0) { if (n1 > 0) {
min = ziamin_kernel_16_TUNED(n1, x, &minf); min = ziamin_kernel_16_TUNED(n1, x, &minf);
i = n1; i = n1;
ix = n1 << 1; ix = n1 << 1;
} }
#endif #endif
while(i < n) while(i < n)
{ {