wernsaar
13 years ago
1 changed files with 8 additions and 0 deletions
Unified View
Diff Options
+ 8
- 0
kernel/x86_64/dtrsm_kernel_LT_8x2_bulldozer.S
View File
| @@ -84,6 +84,9 @@ | |||||
| #endif | #endif | ||||
| #define A_PR1 384 | |||||
| #define B_PR1 192 | |||||
| .macro KERNEL8x2_SUB | .macro KERNEL8x2_SUB | ||||
| vmovddup -16*SIZE(BO,%rax,2), %xmm1 | vmovddup -16*SIZE(BO,%rax,2), %xmm1 | ||||
| @@ -708,9 +711,14 @@ | |||||
| ALIGN_4 | ALIGN_4 | ||||
| .L52: | .L52: | ||||
| prefetcht0 A_PR1(AO,%rax,8) | |||||
| prefetcht0 B_PR1(BO,%rax,2) | |||||
| KERNEL8x2_SUB | KERNEL8x2_SUB | ||||
| prefetcht0 A_PR1(AO,%rax,8) | |||||
| KERNEL8x2_SUB | KERNEL8x2_SUB | ||||
| prefetcht0 A_PR1(AO,%rax,8) | |||||
| KERNEL8x2_SUB | KERNEL8x2_SUB | ||||
| prefetcht0 A_PR1(AO,%rax,8) | |||||
| KERNEL8x2_SUB | KERNEL8x2_SUB | ||||
| jl .L52 | jl .L52 | ||||