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

Merge pull request #2881 from mattip/fninit 

add fninit to reset fpu registers before assembler routines
This commit is contained in:
Martin Kroeker 2020-10-12 23:50:41 +02:00 committed by GitHub
parent cb4274e3ad 403eb513a0
commit 0c84ffe05f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
27 changed files with 107 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
kernel/x86_64/amax.S
View File

@ -54,6 +54,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX

4
kernel/x86_64/asum.S
View File

@ -49,6 +49,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M

4
kernel/x86_64/dot.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX
salq $BASE_SHIFT, INCY salq $BASE_SHIFT, INCY

4
kernel/x86_64/iamax.S
View File

@ -60,6 +60,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX
fldz fldz

4
kernel/x86_64/izamax.S
View File

@ -60,6 +60,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX
fldz fldz

4
kernel/x86_64/nrm2.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

4
kernel/x86_64/qconjg.S
View File

@ -42,6 +42,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
FLD 1 * SIZE(ARG1) FLD 1 * SIZE(ARG1)
fsubrp %st, %st(1) fsubrp %st, %st(1)

4
kernel/x86_64/qdot.S
View File

@ -58,6 +58,10 @@
PROLOGUE PROLOGUE
#ifdef WINDOWS_ABI
emms
#endif
pushl %edi pushl %edi
pushl %esi pushl %esi
pushl %ebx pushl %ebx

4
kernel/x86_64/qgemm_kernel_2x2.S
View File

@ -74,6 +74,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/qgemv_n.S
View File

@ -76,6 +76,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/qgemv_t.S
View File

@ -75,6 +75,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/qtrsm_kernel_LN_2x2.S
View File

@ -74,6 +74,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/qtrsm_kernel_LT_2x2.S
View File

@ -74,6 +74,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/qtrsm_kernel_RT_2x2.S
View File

@ -74,6 +74,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/sum.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

5
kernel/x86_64/xdot.S
View File

@ -59,6 +59,11 @@
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
#define N %ebx #define N %ebx
#define X %esi #define X %esi
#define INCX %ecx #define INCX %ecx

4
kernel/x86_64/xgemm3m_kernel_2x2.S
View File

@ -78,6 +78,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/xgemm_kernel_1x1.S
View File

@ -97,6 +97,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/xgemv_n.S
View File

@ -76,6 +76,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/xgemv_t.S
View File

@ -75,6 +75,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/xtrsm_kernel_LT_1x1.S
View File

@ -90,6 +90,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)
movq %rbp, 8(%rsp) movq %rbp, 8(%rsp)

4
kernel/x86_64/zamax.S
View File

@ -55,6 +55,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX
fldz fldz

4
kernel/x86_64/zasum.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

2
kernel/x86_64/zdot.S
View File

@ -55,6 +55,8 @@
PROFCODE PROFCODE
#ifdef WINDOWS_ABI #ifdef WINDOWS_ABI
emms
movq 40(%rsp), INCY movq 40(%rsp), INCY
#endif #endif

4
kernel/x86_64/znrm2.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

4
kernel/x86_64/zscal.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX
FLD 8(%rsp) FLD 8(%rsp)

4
kernel/x86_64/zsum.S
View File

@ -50,6 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999