moebiusband/MD-Bench
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add comments for second kernel variant on Intel AOS assembly

Browse Source 
Signed-off-by: Rafael Ravedutti <rafaelravedutti@gmail.com>
This commit is contained in:
Rafael Ravedutti 2021-04-16 16:31:27 +02:00
parent e656490a38
commit 5131b7bcaa
1 changed files with 26 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
asm/force_aos_intel.s
View File

@ -573,33 +573,58 @@ computeForce:
# LOE rbx rbp rdi r13 eax edx ecx esi r8d r9d r10d r11d r12d r14d r15d xmm6 xmm7 xmm12 ymm3 ymm15 ymm16 zmm0 zmm1 zmm2 zmm5 zmm8 zmm9 zmm10 zmm11 zmm13 zmm14
..B1.29: # Preds ..B1.26
# Execution count [1.25e+01]
# k1 <- [true for all elements]
vpcmpeqb k1, xmm0, xmm0 #69.36
# k2 <- [true for all elements]
vpcmpeqb k2, xmm0, xmm0 #69.36
# k3 <- [true for all elements]
vpcmpeqb k3, xmm0, xmm0 #69.36
# zmm4 <- 0.0
vpxord zmm4, zmm4, zmm4 #69.36
# zmm17 <- 0.0
vpxord zmm17, zmm17, zmm17 #69.36
# zmm18 <- 0.0
vpxord zmm18, zmm18, zmm18 #69.36
# zmm4 <- atom->x[j * 3 + 2]
vgatherdpd zmm4{k1}, QWORD PTR [16+rdi+ymm3*8] #69.36
# zmm17 <- atom->x[j * 3 + 1]
vgatherdpd zmm17{k2}, QWORD PTR [8+rdi+ymm3*8] #69.36
# zmm18 <- atom->x[j * 3]
vgatherdpd zmm18{k3}, QWORD PTR [rdi+ymm3*8] #69.36
# LOE rbx rbp rdi r13 r11d r12d r14d xmm6 xmm7 xmm12 ymm15 ymm16 zmm0 zmm1 zmm2 zmm4 zmm5 zmm8 zmm9 zmm10 zmm11 zmm13 zmm14 zmm17 zmm18
..B1.30: # Preds ..B1.29
# Execution count [2.50e+01]
# k++
add r12d, 8 #67.9
# k++
add rbx, 8 #67.9
# zmm26 <- atom->x[i * 3 + 2] - atom->x[j * 3 + 2]
vsubpd zmm26, zmm0, zmm4 #71.36
# zmm24 <- atom->x[i * 3 + 1] - atom->x[j * 3 + 1]
vsubpd zmm24, zmm1, zmm17 #70.36
# zmm23 <- atom->x[i * 3] - atom->x[j * 3]
vsubpd zmm23, zmm2, zmm18 #69.36
# zmm3 <- dely * dely
vmulpd zmm3, zmm24, zmm24 #72.49
# zmm3 <- delx * delx + dely * dely
vfmadd231pd zmm3, zmm23, zmm23 #72.49
# zmm3 <- rsq (delz * delz + delx * delx + dely * dely)
vfmadd231pd zmm3, zmm26, zmm26 #72.63
# zmm22 <- sr2 (1.0 / rsq -- compute reciprocal)
vrcp14pd zmm22, zmm3 #75.38
# k2 <- [rsq < cutforcesq]
vcmppd k2, zmm3, zmm14, 1 #74.22
# k0 <- [true when +0, Neg. 0, +inf, -inf]
vfpclasspd k0, zmm22, 30 #75.38
# zmm3 <- [-(rsq * sr2) + 1.0] -- check if 1.0 / rsq is valid! -- call it error?
vfnmadd213pd zmm3, zmm22, QWORD BCST .L_2il0floatpacket.9[rip] #75.38
# k1 <- [false when +0, Neg. 0, +inf, -inf]
knotw k1, k0 #75.38
# zmm4 <- [(-rsq * sr2 + 1.0) * (-rsq * sr2 + 1.0)] -- errorsq
vmulpd zmm4, zmm3, zmm3 #75.38
# zmm22 <- [sr2 * error + sr2]
vfmadd213pd zmm22{k1}, zmm3, zmm22 #75.38
# zmm22 <- [(sr2 * error + sr2) * errorsq + (sr2 * error + sr2)]
vfmadd213pd zmm22{k1}, zmm4, zmm22 #75.38
vmulpd zmm17, zmm22, zmm13 #76.38
vmulpd zmm19, zmm22, zmm10 #77.54
@ -611,6 +636,7 @@ computeForce:
vfmadd231pd zmm9{k2}, zmm25, zmm23 #78.17
vfmadd231pd zmm8{k2}, zmm25, zmm24 #79.17
vfmadd231pd zmm11{k2}, zmm25, zmm26 #80.17
# k < numneighs
cmp r12d, r14d #67.9
jb ..B1.26 # Prob 82% #67.9
# LOE rbx rbp rdi r13 r11d r12d r14d xmm6 xmm7 xmm12 ymm15 ymm16 zmm0 zmm1 zmm2 zmm5 zmm8 zmm9 zmm10 zmm11 zmm13 zmm14