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

Save and restore VSX registers

Browse Source
This commit is contained in:
Martin Kroeker
2017-09-28 12:17:09 +02:00
committed by GitHub
parent 00740c0e34
commit 9c017a2218
15 changed files with 884 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files
kernel/power/dtrsm_kernel_LT_16x4_power8.S
+58 -3
View File
@@ -48,8 +48,9 @@
#ifdef __64BIT__
#define STACKSIZE 320
#define ALPHA 296(SP)
#define FZERO 304(SP)
#define STACKSIZE 520
#define ALPHA 296+200(SP)
#define FZERO 304+200(SP)
#else
#define STACKSIZE 240
#define ALPHA 224(SP)
@@ -112,6 +113,8 @@
#define o48 r30
#define T1 r31
#define VECSAVE r11
#include "dtrsm_macros_LT_16x4_power8.S"
#ifndef NEEDPARAM
@@ -163,6 +166,7 @@
std r17, 256(SP)
std r16, 264(SP)
std r15, 272(SP)
addi r11,SP,288
#else
stw r31, 144(SP)
stw r30, 148(SP)
@@ -178,7 +182,32 @@
stw r20, 188(SP)
stw r19, 192(SP)
stw r18, 196(SP)
addi r11,SP,208
#endif
stvx v20, r11,r0
addi r11,r11,16
stvx v21, r11,r0
addi r11,r11,16
stvx v22, r11,r0
addi r11,r11,16
stvx v23, r11,r0
addi r11,r11,16
stvx v24, r11,r0
addi r11,r11,16
stvx v25, r11,r0
addi r11,r11,16
stvx v26, r11,r0
addi r11,r11,16
stvx v27, r11,r0
addi r11,r11,16
stvx v28, r11,r0
addi r11,r11,16
stvx v29, r11,r0
addi r11,r11,16
stvx v30, r11,r0
addi r11,r11,16
stvx v31, r11,r0
li r11,0
#if defined(_AIX) || defined(__APPLE__)
@@ -269,6 +298,7 @@ L999:
ld r17, 256(SP)
ld r16, 264(SP)
ld r15, 272(SP)
addi r11,SP,288
#else
lwz r31, 144(SP)
lwz r30, 148(SP)
@@ -284,10 +314,35 @@ L999:
lwz r20, 188(SP)
lwz r19, 192(SP)
lwz r18, 196(SP)
addi r11,SP,208
#endif
lvx v20, r11,r3
addi r11,r11,16
lvx v21, r11,r3
addi r11,r11,16
lvx v22, r11,r3
addi r11,r11,16
lvx v23, r11,r3
addi r11,r11,16
lvx v24, r11,r3
addi r11,r11,16
lvx v25, r11,r3
addi r11,r11,16
lvx v26, r11,r3
addi r11,r11,16
lvx v27, r11,r3
addi r11,r11,16
lvx v28, r11,r3
addi r11,r11,16
lvx v29, r11,r3
addi r11,r11,16
lvx v30, r11,r3
addi r11,r11,16
lvx v31, r11,r3
li r11,0
addi SP, SP, STACKSIZE
blr
EPILOGUE