From 95fb98f556adcbbccc5f42318c7c645ec1837e1a Mon Sep 17 00:00:00 2001 From: wjc404 <52632443+wjc404@users.noreply.github.com> Date: Sun, 21 Jul 2019 01:10:32 +0800 Subject: Update dgemm_kernel_4x8_haswell.S --- kernel/x86_64/dgemm_kernel_4x8_haswell.S | 14 +++++++------- 1 file 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 -- cgit v1.2.3