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

modified dtrsm_kernel_LT_8x2_bulldozer.S

Browse Source
This commit is contained in:
wernsaar
2013-08-04 12:16:12 +02:00
parent 8aeec32ea0
commit aaeb8eaecd
1 changed files with 23 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

49
kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S
View File

@@ -310,7 +310,6 @@
vmovddup -1 * SIZE(AO), %xmm8
vmulpd %xmm3 , %xmm8 , %xmm3
vmovsd %xmm0 , 0 * SIZE(CO1)
vmovsd %xmm1 , 1 * SIZE(CO1)
vmovsd %xmm2 , 2 * SIZE(CO1)
@@ -359,7 +358,6 @@
vmovddup -13 * SIZE(AO), %xmm8
vmulpd %xmm1 , %xmm8 , %xmm1
vmovsd %xmm0 , 0 * SIZE(CO1)
vmovsd %xmm1 , 1 * SIZE(CO1)
@@ -481,7 +479,6 @@
vmulsd 47 * SIZE(AO), %xmm7 , %xmm7
vmovsd %xmm0 , 0 * SIZE(CO1)
vmovsd %xmm1 , 1 * SIZE(CO1)
vmovsd %xmm2 , 2 * SIZE(CO1)
@@ -706,7 +703,7 @@
leaq (AO, %rax, 8), AO
leaq (BO, %rax, 2), BO
negq %rax
NOBRANCH
je .L56
ALIGN_4
@@ -765,10 +762,10 @@
movq B, BO
pxor %xmm8, %xmm8
pxor %xmm9, %xmm9
pxor %xmm10, %xmm10
pxor %xmm11, %xmm11
vxorpd %xmm8 , %xmm8 , %xmm8
vxorpd %xmm9 , %xmm9 , %xmm9
vxorpd %xmm10, %xmm10, %xmm10
vxorpd %xmm11, %xmm11, %xmm11
movq KK, %rax
andq $-4, %rax
@@ -776,7 +773,7 @@
leaq (AO, %rax, 4), AO
leaq (BO, %rax, 2), BO
negq %rax
NOBRANCH
je .L56_A
ALIGN_4
@@ -835,8 +832,8 @@
.L61:
movq B, BO
pxor %xmm8, %xmm8
pxor %xmm9, %xmm9
vxorpd %xmm8, %xmm8 , %xmm8
vxorpd %xmm9, %xmm9 , %xmm9
movq KK, %rax
andq $-4, %rax
@@ -844,7 +841,7 @@
leaq (AO, %rax, 2), AO
leaq (BO, %rax, 2), BO
negq %rax
NOBRANCH
je .L66
ALIGN_4
@@ -901,7 +898,7 @@
.L71:
movq B, BO
pxor %xmm8, %xmm8
vxorpd %xmm8, %xmm8 , %xmm8
movq KK, %rax
andq $-4, %rax
@@ -909,7 +906,7 @@
leaq (AO, %rax, 1), AO
leaq (BO, %rax, 2), BO
negq %rax
NOBRANCH
je .L76
ALIGN_4
@@ -985,10 +982,10 @@
movq B, BO
pxor %xmm8, %xmm8
pxor %xmm9, %xmm9
pxor %xmm10, %xmm10
pxor %xmm11, %xmm11
vxorpd %xmm8, %xmm8 , %xmm8
vxorpd %xmm9, %xmm9 , %xmm9
vxorpd %xmm10, %xmm10, %xmm10
vxorpd %xmm11, %xmm11, %xmm11
movq KK, %rax
@@ -997,7 +994,7 @@
leaq (AO, %rax, 8), AO
leaq (BO, %rax, 1), BO
negq %rax
NOBRANCH
je .L96
ALIGN_4
@@ -1053,8 +1050,8 @@
.L91_A:
movq B, BO
pxor %xmm8, %xmm8
pxor %xmm9, %xmm9
vxorpd %xmm8, %xmm8 , %xmm8
vxorpd %xmm9, %xmm9 , %xmm9
movq KK, %rax
andq $-4, %rax
@@ -1062,7 +1059,7 @@
leaq (AO, %rax, 4), AO
leaq (BO, %rax, 1), BO
negq %rax
NOBRANCH
je .L96_A
ALIGN_4
@@ -1119,7 +1116,7 @@
movq B, BO
pxor %xmm8, %xmm8
vxorpd %xmm8, %xmm8 , %xmm8
movq KK, %rax
andq $-4, %rax
@@ -1127,7 +1124,7 @@
leaq (AO, %rax, 2), AO
leaq (BO, %rax, 1), BO
negq %rax
NOBRANCH
je .L106
ALIGN_4
@@ -1183,7 +1180,7 @@
.L111:
movq B, BO
pxor %xmm8, %xmm8
vxorpd %xmm8, %xmm8 , %xmm8
movq KK, %rax
andq $-4, %rax
@@ -1191,7 +1188,7 @@
leaq (AO, %rax, 1), AO
leaq (BO, %rax, 1), BO
negq %rax
NOBRANCH
je .L116
ALIGN_4