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

69
kernel/power/strmm_kernel_16x8_power8.S
View File

@@ -83,8 +83,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#ifdef __64BIT__
#define STACKSIZE 340
#define ALPHA_SP 296(SP)
#define FZERO 304(SP)
#define STACKSIZE 540
#define ALPHA_SP 296+200(SP)
#define FZERO 304+200(SP)
#else
#define STACKSIZE 240
#define ALPHA_SP 224(SP)
@@ -132,6 +133,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define o0 0
#define VECSAVE r11
#define TBUFFER r13
#define o12 r14
#define o4 r15
@@ -207,6 +210,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
std r15, 272(SP)
std r14, 280(SP)
std r13, 288(SP)
addi r11, SP, 304
#else
stw r31, 144(SP)
stw r30, 148(SP)
@@ -226,8 +230,34 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
stw r16, 204(SP)
stw r15, 208(SP)
stw r14, 212(SP)
stw r13, 216(SP)
stw r13, 216(SP)
addi r11, SP, 224
#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
// stfd f1, ALPHA_SP
// stw r0, FZERO
@@ -271,16 +301,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
cmpwi cr0, K, 0
ble L999_H1
li PRE, 256
li PRE, 256
li o4 , 4
li o8 , 8
li o12, 12
li o16, 16
li o32, 32
li o48, 48
addi TBUFFER, SP, 320
addi TBUFFER, SP, 320+200
addi T1, SP, 300
addi T1, SP, 300+200
stxsspx f1, o0 , T1
stxsspx f1, o4 , T1
stxsspx f1, o8 , T1
@@ -339,6 +369,7 @@ L999:
ld r15, 272(SP)
ld r14, 280(SP)
ld r13, 288(SP)
addi r11, SP, 304
#else
lwz r31, 144(SP)
lwz r30, 148(SP)
@@ -359,10 +390,34 @@ L999:
lwz r15, 208(SP)
lwz r14, 212(SP)
lwz r13, 216(SP)
addi r11, SP, 224
#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