@@ -267,24 +267,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
267
267
268
268
.macro SAVE4x12
269
269
270
- prefetcht0 128 ( %rsp ) /*BUFFER 1*/
270
+ prefetcht0 BUFFER1
271
271
vbroadcastsd ALPHA, %ymm0
272
272
273
273
vmulpd %ymm0 , %ymm4 , %ymm4
274
274
vmulpd %ymm0 , %ymm5 , %ymm5
275
275
vmulpd %ymm0 , %ymm6 , %ymm6
276
276
vmulpd %ymm0 , %ymm7 , %ymm7
277
- prefetcht0 192 ( %rsp )
277
+ prefetcht0 64 + BUFFER1
278
278
vmulpd %ymm0 , %ymm8 , %ymm8
279
279
vmulpd %ymm0 , %ymm9 , %ymm9
280
280
vmulpd %ymm0 , %ymm10 , %ymm10
281
281
vmulpd %ymm0 , %ymm11 , %ymm11
282
- prefetcht0 256 ( %rsp )
282
+ prefetcht0 128 + BUFFER1
283
283
vmulpd %ymm0 , %ymm12 , %ymm12
284
284
vmulpd %ymm0 , %ymm13 , %ymm13
285
285
vmulpd %ymm0 , %ymm14 , %ymm14
286
286
vmulpd %ymm0 , %ymm15 , %ymm15
287
- prefetcht0 320 ( %rsp )
287
+ prefetcht0 192 + BUFFER1
288
288
vpermilpd $ 0x05 , %ymm5 , %ymm5
289
289
vpermilpd $ 0x05 , %ymm7 , %ymm7
290
290
@@ -1606,6 +1606,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1606
1606
.endm
1607
1607
1608
1608
1609
+ .macro PREFETCHT0_C
1610
+ prefetcht0 (CO1)
1611
+ prefetcht0 24 (CO1)
1612
+ prefetcht0 (CO1,LDC,4 )
1613
+ prefetcht0 24 (CO1,LDC,4 )
1614
+ prefetcht0 (CO1,LDC,8 )
1615
+ prefetcht0 24 (CO1,LDC,8 )
1616
+ addq LDC,CO1
1617
+ prefetcht0 (CO1)
1618
+ prefetcht0 24 (CO1)
1619
+ prefetcht0 (CO1,LDC,4 )
1620
+ prefetcht0 24 (CO1,LDC,4 )
1621
+ prefetcht0 (CO1,LDC,8 )
1622
+ prefetcht0 24 (CO1,LDC,8 )
1623
+ leaq (CO1,LDC,2 ),CO1
1624
+ prefetcht0 (CO1)
1625
+ prefetcht0 24 (CO1)
1626
+ prefetcht0 (CO1,LDC,4 )
1627
+ prefetcht0 24 (CO1,LDC,4 )
1628
+ prefetcht0 (CO1,LDC,8 )
1629
+ prefetcht0 24 (CO1,LDC,8 )
1630
+ subq LDC,CO1
1631
+ prefetcht0 (CO1)
1632
+ prefetcht0 24 (CO1)
1633
+ prefetcht0 (CO1,LDC,4 )
1634
+ prefetcht0 24 (CO1,LDC,4 )
1635
+ prefetcht0 (CO1,LDC,8 )
1636
+ prefetcht0 24 (CO1,LDC,8 )
1637
+ subq LDC,CO1
1638
+ subq LDC,CO1
1639
+ .endm
1609
1640
/*******************************************************************************************/
1610
1641
1611
1642
#if !defined(TRMMKERNEL)
@@ -1773,7 +1804,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
1773
1804
1774
1805
dec %rax
1775
1806
jne .L12_12
1776
-
1807
+
1808
+ PREFETCHT0_C
1777
1809
.L12_12a:
1778
1810
1779
1811
KERNEL4x12_M1
0 commit comments