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

use emms instead, add WIN guards

This commit is contained in:
Matti Picus 2020-10-12 18:15:01 +03:00
parent a5b164946c
commit 403eb513a0
27 changed files with 87 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
kernel/x86_64/amax.S
View File

@ -55,7 +55,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX

5
kernel/x86_64/asum.S
View File

@ -50,7 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

5
kernel/x86_64/dot.S
View File

@ -49,7 +49,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit
#ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX
salq $BASE_SHIFT, INCY salq $BASE_SHIFT, INCY

5
kernel/x86_64/iamax.S
View File

@ -59,7 +59,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit
#ifdef WINDOWS_ABI
emms
#endif
salq $BASE_SHIFT, INCX salq $BASE_SHIFT, INCX

5
kernel/x86_64/izamax.S
View File

@ -59,7 +59,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit
#ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX

5
kernel/x86_64/nrm2.S
View File

@ -50,7 +50,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M
jle .L999 jle .L999

5
kernel/x86_64/qconjg.S
View File

@ -41,7 +41,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit
#ifdef WINDOWS_ABI
emms
#endif
fldz fldz
FLD 1 * SIZE(ARG1) FLD 1 * SIZE(ARG1)

4
kernel/x86_64/qdot.S
View File

@ -58,7 +58,9 @@
PROLOGUE PROLOGUE
fninit #ifdef WINDOWS_ABI
emms
#endif
pushl %edi pushl %edi
pushl %esi pushl %esi

4
kernel/x86_64/qgemm_kernel_2x2.S
View File

@ -74,7 +74,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/qgemv_n.S
View File

@ -76,7 +76,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

5
kernel/x86_64/qgemv_t.S
View File

@ -75,7 +75,10 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #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

@ -74,7 +74,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/qtrsm_kernel_LT_2x2.S
View File

@ -74,7 +74,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

5
kernel/x86_64/qtrsm_kernel_RT_2x2.S
View File

@ -74,8 +74,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/sum.S
View File

@ -50,7 +50,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M

4
kernel/x86_64/xdot.S
View File

@ -59,7 +59,9 @@
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
#define N %ebx #define N %ebx

4
kernel/x86_64/xgemm3m_kernel_2x2.S
View File

@ -78,7 +78,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/xgemm_kernel_1x1.S
View File

@ -97,7 +97,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/xgemv_n.S
View File

@ -76,7 +76,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/xgemv_t.S
View File

@ -75,7 +75,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/xtrsm_kernel_LT_1x1.S
View File

@ -90,7 +90,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
subq $STACKSIZE, %rsp subq $STACKSIZE, %rsp
movq %rbx, 0(%rsp) movq %rbx, 0(%rsp)

4
kernel/x86_64/zamax.S
View File

@ -55,7 +55,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX

4
kernel/x86_64/zasum.S
View File

@ -50,7 +50,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M

4
kernel/x86_64/zdot.S
View File

@ -54,9 +54,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit
#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

@ -50,7 +50,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M

4
kernel/x86_64/zscal.S
View File

@ -50,7 +50,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
salq $ZBASE_SHIFT, INCX salq $ZBASE_SHIFT, INCX

4
kernel/x86_64/zsum.S
View File

@ -50,7 +50,9 @@
PROLOGUE PROLOGUE
PROFCODE PROFCODE
fninit #ifdef WINDOWS_ABI
emms
#endif
fldz fldz
testq M, M testq M, M