diff options
| author | wjc404 <52632443+wjc404@users.noreply.github.com> | 2019-07-21 01:10:32 +0800 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2019-07-21 01:10:32 +0800 |
| commit | 95fb98f556adcbbccc5f42318c7c645ec1837e1a (patch) | |
| tree | 5897ffe71bfc206d668d05ef93ca189f2856d05b | |
| parent | 4801c6d36bd87421b08e60efa1b6e0217fd41672 (diff) | |
| download | openblas-95fb98f556adcbbccc5f42318c7c645ec1837e1a.tar.gz openblas-95fb98f556adcbbccc5f42318c7c645ec1837e1a.tar.bz2 openblas-95fb98f556adcbbccc5f42318c7c645ec1837e1a.zip | |
Update dgemm_kernel_4x8_haswell.S
| -rw-r--r-- | kernel/x86_64/dgemm_kernel_4x8_haswell.S | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/kernel/x86_64/dgemm_kernel_4x8_haswell.S b/kernel/x86_64/dgemm_kernel_4x8_haswell.S index 26eea0acf..082e62a7c 100644 --- a/kernel/x86_64/dgemm_kernel_4x8_haswell.S +++ b/kernel/x86_64/dgemm_kernel_4x8_haswell.S @@ -279,43 +279,43 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. vmulpd %ymm0 , %ymm9 , %ymm9
vmulpd %ymm0 , %ymm10, %ymm10
vmulpd %ymm0 , %ymm11, %ymm11
-#if B_PR1 >= 96
+#if B_PR1 > 32
prefetcht0 128 + BUFFER1
#endif
vmulpd %ymm0 , %ymm12, %ymm12
vmulpd %ymm0 , %ymm13, %ymm13
vmulpd %ymm0 , %ymm14, %ymm14
vmulpd %ymm0 , %ymm15, %ymm15
-#if B_PR1 >= 160
+#if B_PR1 > 96
prefetcht0 192 + BUFFER1
#endif
vpermilpd $ 0x05 , %ymm5, %ymm5
vpermilpd $ 0x05 , %ymm7, %ymm7
-#if B_PR1 >= 224
+#if B_PR1 > 160
prefetcht0 256 + BUFFER1
#endif
vblendpd $ 0x0a, %ymm5, %ymm4, %ymm0
vblendpd $ 0x05, %ymm5, %ymm4, %ymm1
vblendpd $ 0x0a, %ymm7, %ymm6, %ymm2
vblendpd $ 0x05, %ymm7, %ymm6, %ymm3
-#if B_PR1 >= 288
+#if B_PR1 > 224
prefetcht0 320 + BUFFER1
#endif
vperm2f128 $ 0x01 , %ymm2, %ymm2 , %ymm2
vperm2f128 $ 0x01 , %ymm3, %ymm3 , %ymm3
-#if B_PR1 >= 352
+#if B_PR1 > 288
prefetcht0 384 + BUFFER1
#endif
vblendpd $ 0x03, %ymm0, %ymm2 , %ymm4
vblendpd $ 0x03, %ymm1, %ymm3 , %ymm5
vblendpd $ 0x03, %ymm2, %ymm0 , %ymm6
vblendpd $ 0x03, %ymm3, %ymm1 , %ymm7
-#if B_PR1 >= 416
+#if B_PR1 > 352
prefetcht0 448 + BUFFER1
#endif
leaq (CO1, LDC, 2), %rax
-#if B_PR1 >= 480
+#if B_PR1 > 416
prefetcht0 512 + BUFFER1
#endif
|