modified dtrsm_kernel_LT_8x2_bulldozer.S

diff --git a/kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S b/kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S
index bc15537..e9b3665 100644 (file)
--- a/kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S
+++ b/kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S
@@ -84,12 +84,6 @@
 
 #endif
 
-#define PREFETCH     prefetch
-#define PREFETCHSIZE  (8 *  7 + 0)
-
-#define movlpd vmovsd
-#define movapd vmovups
-#define movupd vmovups
 
 .macro KERNEL8x2_SUB
        vmovddup        -16*SIZE(BO,%rax,2), %xmm1
@@ -97,15 +91,15 @@
        vmovups         -16*SIZE(AO,%rax,8), %xmm0
        vfmaddpd        %xmm8 , %xmm0 , %xmm1 , %xmm8
        vfmaddpd        %xmm9 , %xmm0 , %xmm2 , %xmm9
-       vmovups         -14*SIZE(AO,%rax,8), %xmm0
-       vfmaddpd        %xmm10, %xmm0 , %xmm1 , %xmm10
-       vfmaddpd        %xmm11, %xmm0 , %xmm2 , %xmm11
-       vmovups         -12*SIZE(AO,%rax,8), %xmm0
-       vfmaddpd        %xmm12, %xmm0 , %xmm1 , %xmm12
-       vfmaddpd        %xmm13, %xmm0 , %xmm2 , %xmm13
-       vmovups         -10*SIZE(AO,%rax,8), %xmm0
-       vfmaddpd        %xmm14, %xmm0 , %xmm1 , %xmm14
-       vfmaddpd        %xmm15, %xmm0 , %xmm2 , %xmm15
+       vmovups         -14*SIZE(AO,%rax,8), %xmm4
+       vfmaddpd        %xmm10, %xmm4 , %xmm1 , %xmm10
+       vfmaddpd        %xmm11, %xmm4 , %xmm2 , %xmm11
+       vmovups         -12*SIZE(AO,%rax,8), %xmm5
+       vfmaddpd        %xmm12, %xmm5 , %xmm1 , %xmm12
+       vfmaddpd        %xmm13, %xmm5 , %xmm2 , %xmm13
+       vmovups         -10*SIZE(AO,%rax,8), %xmm6
+       vfmaddpd        %xmm14, %xmm6 , %xmm1 , %xmm14
+       vfmaddpd        %xmm15, %xmm6 , %xmm2 , %xmm15
        addq    $SIZE, %rax
 .endm