floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update dgemm_kernel_4x8_haswell.S

This commit is contained in:
wjc404 2019-07-21 01:10:32 +08:00 committed by GitHub
parent 4801c6d36b
commit 95fb98f556
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
kernel/x86_64/dgemm_kernel_4x8_haswell.S
View File

@ -279,43 +279,43 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
vmulpd %ymm0 , %ymm9 , %ymm9 vmulpd %ymm0 , %ymm9 , %ymm9
vmulpd %ymm0 , %ymm10, %ymm10 vmulpd %ymm0 , %ymm10, %ymm10
vmulpd %ymm0 , %ymm11, %ymm11 vmulpd %ymm0 , %ymm11, %ymm11
#if B_PR1 >= 96 #if B_PR1 > 32
prefetcht0 128 + BUFFER1 prefetcht0 128 + BUFFER1
#endif #endif
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
#if B_PR1 >= 160 #if B_PR1 > 96
prefetcht0 192 + BUFFER1 prefetcht0 192 + BUFFER1
#endif #endif
vpermilpd $ 0x05 , %ymm5, %ymm5 vpermilpd $ 0x05 , %ymm5, %ymm5
vpermilpd $ 0x05 , %ymm7, %ymm7 vpermilpd $ 0x05 , %ymm7, %ymm7
#if B_PR1 >= 224 #if B_PR1 > 160
prefetcht0 256 + BUFFER1 prefetcht0 256 + BUFFER1
#endif #endif
vblendpd $ 0x0a, %ymm5, %ymm4, %ymm0 vblendpd $ 0x0a, %ymm5, %ymm4, %ymm0
vblendpd $ 0x05, %ymm5, %ymm4, %ymm1 vblendpd $ 0x05, %ymm5, %ymm4, %ymm1
vblendpd $ 0x0a, %ymm7, %ymm6, %ymm2 vblendpd $ 0x0a, %ymm7, %ymm6, %ymm2
vblendpd $ 0x05, %ymm7, %ymm6, %ymm3 vblendpd $ 0x05, %ymm7, %ymm6, %ymm3
#if B_PR1 >= 288 #if B_PR1 > 224
prefetcht0 320 + BUFFER1 prefetcht0 320 + BUFFER1
#endif #endif
vperm2f128 $ 0x01 , %ymm2, %ymm2 , %ymm2 vperm2f128 $ 0x01 , %ymm2, %ymm2 , %ymm2
vperm2f128 $ 0x01 , %ymm3, %ymm3 , %ymm3 vperm2f128 $ 0x01 , %ymm3, %ymm3 , %ymm3
#if B_PR1 >= 352 #if B_PR1 > 288
prefetcht0 384 + BUFFER1 prefetcht0 384 + BUFFER1
#endif #endif
vblendpd $ 0x03, %ymm0, %ymm2 , %ymm4 vblendpd $ 0x03, %ymm0, %ymm2 , %ymm4
vblendpd $ 0x03, %ymm1, %ymm3 , %ymm5 vblendpd $ 0x03, %ymm1, %ymm3 , %ymm5
vblendpd $ 0x03, %ymm2, %ymm0 , %ymm6 vblendpd $ 0x03, %ymm2, %ymm0 , %ymm6
vblendpd $ 0x03, %ymm3, %ymm1 , %ymm7 vblendpd $ 0x03, %ymm3, %ymm1 , %ymm7
#if B_PR1 >= 416 #if B_PR1 > 352
prefetcht0 448 + BUFFER1 prefetcht0 448 + BUFFER1
#endif #endif
leaq (CO1, LDC, 2), %rax leaq (CO1, LDC, 2), %rax
#if B_PR1 >= 480 #if B_PR1 > 416
prefetcht0 512 + BUFFER1 prefetcht0 512 + BUFFER1
#endif #endif