| @@ -267,24 +267,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |||||
| .macro SAVE4x12 | .macro SAVE4x12 | ||||
| prefetcht0 128(%rsp) /*BUFFER 1*/ | |||||
| prefetcht0 BUFFER1 | |||||
| vbroadcastsd ALPHA, %ymm0 | vbroadcastsd ALPHA, %ymm0 | ||||
| vmulpd %ymm0 , %ymm4 , %ymm4 | vmulpd %ymm0 , %ymm4 , %ymm4 | ||||
| vmulpd %ymm0 , %ymm5 , %ymm5 | vmulpd %ymm0 , %ymm5 , %ymm5 | ||||
| vmulpd %ymm0 , %ymm6 , %ymm6 | vmulpd %ymm0 , %ymm6 , %ymm6 | ||||
| vmulpd %ymm0 , %ymm7 , %ymm7 | vmulpd %ymm0 , %ymm7 , %ymm7 | ||||
| prefetcht0 192(%rsp) | |||||
| prefetcht0 64 + BUFFER1 | |||||
| vmulpd %ymm0 , %ymm8 , %ymm8 | vmulpd %ymm0 , %ymm8 , %ymm8 | ||||
| vmulpd %ymm0 , %ymm9 , %ymm9 | vmulpd %ymm0 , %ymm9 , %ymm9 | ||||
| vmulpd %ymm0 , %ymm10, %ymm10 | vmulpd %ymm0 , %ymm10, %ymm10 | ||||
| vmulpd %ymm0 , %ymm11, %ymm11 | vmulpd %ymm0 , %ymm11, %ymm11 | ||||
| prefetcht0 256(%rsp) | |||||
| prefetcht0 128 + BUFFER1 | |||||
| vmulpd %ymm0 , %ymm12, %ymm12 | vmulpd %ymm0 , %ymm12, %ymm12 | ||||
| vmulpd %ymm0 , %ymm13, %ymm13 | vmulpd %ymm0 , %ymm13, %ymm13 | ||||
| vmulpd %ymm0 , %ymm14, %ymm14 | vmulpd %ymm0 , %ymm14, %ymm14 | ||||
| vmulpd %ymm0 , %ymm15, %ymm15 | vmulpd %ymm0 , %ymm15, %ymm15 | ||||
| prefetcht0 320(%rsp) | |||||
| prefetcht0 192 + BUFFER1 | |||||
| vpermilpd $ 0x05 , %ymm5, %ymm5 | vpermilpd $ 0x05 , %ymm5, %ymm5 | ||||
| vpermilpd $ 0x05 , %ymm7, %ymm7 | vpermilpd $ 0x05 , %ymm7, %ymm7 | ||||
| @@ -1606,6 +1606,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |||||
| .endm | .endm | ||||
| .macro PREFETCHT0_C | |||||
| prefetcht0 (CO1) | |||||
| prefetcht0 24(CO1) | |||||
| prefetcht0 (CO1,LDC,4) | |||||
| prefetcht0 24(CO1,LDC,4) | |||||
| prefetcht0 (CO1,LDC,8) | |||||
| prefetcht0 24(CO1,LDC,8) | |||||
| addq LDC,CO1 | |||||
| prefetcht0 (CO1) | |||||
| prefetcht0 24(CO1) | |||||
| prefetcht0 (CO1,LDC,4) | |||||
| prefetcht0 24(CO1,LDC,4) | |||||
| prefetcht0 (CO1,LDC,8) | |||||
| prefetcht0 24(CO1,LDC,8) | |||||
| leaq (CO1,LDC,2),CO1 | |||||
| prefetcht0 (CO1) | |||||
| prefetcht0 24(CO1) | |||||
| prefetcht0 (CO1,LDC,4) | |||||
| prefetcht0 24(CO1,LDC,4) | |||||
| prefetcht0 (CO1,LDC,8) | |||||
| prefetcht0 24(CO1,LDC,8) | |||||
| subq LDC,CO1 | |||||
| prefetcht0 (CO1) | |||||
| prefetcht0 24(CO1) | |||||
| prefetcht0 (CO1,LDC,4) | |||||
| prefetcht0 24(CO1,LDC,4) | |||||
| prefetcht0 (CO1,LDC,8) | |||||
| prefetcht0 24(CO1,LDC,8) | |||||
| subq LDC,CO1 | |||||
| subq LDC,CO1 | |||||
| .endm | |||||
| /*******************************************************************************************/ | /*******************************************************************************************/ | ||||
| #if !defined(TRMMKERNEL) | #if !defined(TRMMKERNEL) | ||||
| @@ -1773,7 +1804,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |||||
| dec %rax | dec %rax | ||||
| jne .L12_12 | jne .L12_12 | ||||
| PREFETCHT0_C | |||||
| .L12_12a: | .L12_12a: | ||||
| KERNEL4x12_M1 | KERNEL4x12_M1 | ||||