Skip to content

Commit 94db259

Browse files
authored
Add files via upload
1 parent f49f804 commit 94db259

File tree

1 file changed

+19
-26
lines changed

1 file changed

+19
-26
lines changed

kernel/x86_64/dgemm_kernel_4x8_haswell.S

Lines changed: 19 additions & 26 deletions
Original file line numberDiff line numberDiff line change
@@ -1622,35 +1622,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
16221622

16231623

16241624
.macro PREFETCHT0_C
1625+
prefetcht0 ALPHA
16251626
prefetcht0 (CO1)
16261627
prefetcht0 24(CO1)
16271628
prefetcht0 (CO1,LDC,4)
16281629
prefetcht0 24(CO1,LDC,4)
16291630
prefetcht0 (CO1,LDC,8)
16301631
prefetcht0 24(CO1,LDC,8)
1631-
addq LDC,CO1
1632-
prefetcht0 (CO1)
1633-
prefetcht0 24(CO1)
1634-
prefetcht0 (CO1,LDC,4)
1635-
prefetcht0 24(CO1,LDC,4)
1636-
prefetcht0 (CO1,LDC,8)
1637-
prefetcht0 24(CO1,LDC,8)
1638-
leaq (CO1,LDC,2),CO1
1639-
prefetcht0 (CO1)
1640-
prefetcht0 24(CO1)
1641-
prefetcht0 (CO1,LDC,4)
1642-
prefetcht0 24(CO1,LDC,4)
1643-
prefetcht0 (CO1,LDC,8)
1644-
prefetcht0 24(CO1,LDC,8)
1645-
subq LDC,CO1
1646-
prefetcht0 (CO1)
1647-
prefetcht0 24(CO1)
1648-
prefetcht0 (CO1,LDC,4)
1649-
prefetcht0 24(CO1,LDC,4)
1650-
prefetcht0 (CO1,LDC,8)
1651-
prefetcht0 24(CO1,LDC,8)
1652-
subq LDC,CO1
1653-
subq LDC,CO1
16541632
.endm
16551633
/*******************************************************************************************/
16561634

@@ -1820,12 +1798,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
18201798
dec %rax
18211799
jne .L12_12
18221800

1823-
PREFETCHT0_C
18241801
.L12_12a:
1825-
1802+
PREFETCHT0_C
1803+
addq LDC,CO1
18261804
KERNEL4x12_M1
1805+
PREFETCHT0_C
1806+
leaq (CO1,LDC,2),CO1
18271807
KERNEL4x12_M2
1808+
PREFETCHT0_C
1809+
subq LDC,CO1
18281810
KERNEL4x12_M1
1811+
PREFETCHT0_C
1812+
subq LDC,CO1
1813+
subq LDC,CO1
18291814
KERNEL4x12_M2
18301815

18311816
KERNEL4x12_M1
@@ -2133,9 +2118,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
21332118

21342119
.L13_12a:
21352120

2121+
PREFETCHT0_C
2122+
addq LDC,CO1
21362123
KERNEL4x12_M1
2124+
PREFETCHT0_C
2125+
leaq (CO1,LDC,2),CO1
21372126
KERNEL4x12_M2
2127+
PREFETCHT0_C
2128+
subq LDC,CO1
21382129
KERNEL4x12_M1
2130+
PREFETCHT0_C
2131+
subq LDC,CO1
2132+
subq LDC,CO1
21392133
KERNEL4x12_M2
21402134

21412135
KERNEL4x12_M1
@@ -2145,7 +2139,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
21452139

21462140
jmp .L13_16
21472141

2148-
PREFETCHT0_C
21492142
.L13_13:
21502143

21512144
test $1, %rax

0 commit comments

Comments
 (0)