nb_kernel310_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,781 行 · 第 1/4 页
S
1,781 行
addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm7,%xmm7 ## 2*Heps2 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps mulps %xmm3,%xmm5 ## eps*Fp addps %xmm4,%xmm5 ## VV mulps nb310_qq(%rsp),%xmm5 ## VV*qq=vcoul mulps nb310_qq(%rsp),%xmm7 ## FF*qq=fijC ## LJ forces mulps nb310_six(%rsp),%xmm2 mulps nb310_twelve(%rsp),%xmm0 subps %xmm2,%xmm0 mulps %xmm1,%xmm0 ## (12*vnb12-6*vnb6)*rinv ## add potential to vctot addps nb310_vctot(%rsp),%xmm5 movaps %xmm5,nb310_vctot(%rsp) mulps nb310_tsc(%rsp),%xmm7 subps %xmm7,%xmm0 mulps %xmm1,%xmm0 ## fscal ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm0,%xmm9 mulps %xmm0,%xmm10 mulps %xmm0,%xmm11 movq nb310_faction(%rbp),%rsi ## the fj's - start by accumulating x & y forces from memory movlps (%rsi,%rax,4),%xmm0 ## x1 y1 - - movlps (%rsi,%rcx,4),%xmm1 ## x3 y3 - - movhps (%rsi,%rbx,4),%xmm0 ## x1 y1 x2 y2 movhps (%rsi,%rdx,4),%xmm1 ## x3 y3 x4 y4 ## xmm0-xmm2 contains tx-tz (partial force) ## accumulate i forces addps %xmm9,%xmm13 addps %xmm10,%xmm14 addps %xmm11,%xmm15 movaps %xmm9,%xmm8 unpcklps %xmm10,%xmm9 ## x1 y1 x2 y2 unpckhps %xmm10,%xmm8 ## x3 y3 x4 y4 ## update fjx and fjy addps %xmm9,%xmm0 addps %xmm8,%xmm1 movlps %xmm0,(%rsi,%rax,4) movlps %xmm1,(%rsi,%rcx,4) movhps %xmm0,(%rsi,%rbx,4) movhps %xmm1,(%rsi,%rdx,4) ## xmm11: fjz1 fjz2 fjz3 fjz4 pshufd $1,%xmm11,%xmm10 ## fjz2 - - - movhlps %xmm11,%xmm9 ## fjz3 - - - pshufd $3,%xmm11,%xmm8 ## fjz4 - - - addss 8(%rsi,%rax,4),%xmm11 addss 8(%rsi,%rbx,4),%xmm10 addss 8(%rsi,%rcx,4),%xmm9 addss 8(%rsi,%rdx,4),%xmm8 movss %xmm11,8(%rsi,%rax,4) movss %xmm10,8(%rsi,%rbx,4) movss %xmm9,8(%rsi,%rcx,4) movss %xmm8,8(%rsi,%rdx,4) ## should we do one more iteration? subl $4,nb310_innerk(%rsp) jl _nb_kernel310_x86_64_sse.nb310_finish_inner jmp _nb_kernel310_x86_64_sse.nb310_unroll_loop_nb_kernel310_x86_64_sse.nb310_finish_inner: ## check if at least two particles remain addl $4,nb310_innerk(%rsp) movl nb310_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel310_x86_64_sse.nb310_dopair jmp _nb_kernel310_x86_64_sse.nb310_checksingle_nb_kernel310_x86_64_sse.nb310_dopair: ## twice-unrolled innerloop here movq nb310_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb310_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb310_charge(%rbp),%rsi movss (%rsi,%rax,4),%xmm0 movss (%rsi,%rbx,4),%xmm2 unpcklps %xmm2,%xmm0 ## jqa jqb mulps nb310_iq(%rsp),%xmm0 movaps %xmm0,nb310_qq(%rsp) movq nb310_type(%rbp),%rsi ## vdw parameters movl (%rsi,%rax,4),%r12d movl (%rsi,%rbx,4),%r13d shll %r12d shll %r13d movl nb310_ntia(%rsp),%edi addl %edi,%r12d addl %edi,%r13d movq nb310_vdwparam(%rbp),%rsi movlps (%rsi,%r12,4),%xmm3 movhps (%rsi,%r13,4),%xmm3 xorps %xmm7,%xmm7 movaps %xmm3,%xmm0 shufps $136,%xmm7,%xmm0 ## 10001000 shufps $221,%xmm7,%xmm3 ## 11011101 movaps %xmm0,nb310_c6(%rsp) movaps %xmm3,nb310_c12(%rsp) lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## load coordinates movq nb310_pos(%rbp),%rdi movlps (%rdi,%rax,4),%xmm4 ## x1 y1 - - movlps (%rdi,%rbx,4),%xmm5 ## x2 y2 - - movss 8(%rdi,%rax,4),%xmm6 ## z1 - - - movss 8(%rdi,%rbx,4),%xmm7 ## z2 - - - unpcklps %xmm5,%xmm4 ## x1 x2 y1 y2 movhlps %xmm4,%xmm5 ## y1 y2 - - unpcklps %xmm7,%xmm6 ## z1 z2 - - ## calc dr subps nb310_ix(%rsp),%xmm4 subps nb310_iy(%rsp),%xmm5 subps nb310_iz(%rsp),%xmm6 ## store dr in xmm9-xmm11 movaps %xmm4,%xmm9 movaps %xmm5,%xmm10 movaps %xmm6,%xmm11 ## square it mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 mulps %xmm6,%xmm6 addps %xmm5,%xmm4 addps %xmm6,%xmm4 ## rsq in xmm4 ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm4,%xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb310_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps nb310_half(%rsp),%xmm1 ## xmm1=rinv movaps %xmm4,%xmm3 ## xmm3=rsq mulps %xmm1,%xmm3 ## r mulps nb310_tsc(%rsp),%xmm3 ## rtab ## truncate and convert to integers cvttps2dq %xmm3,%xmm2 ## convert back to float cvtdq2ps %xmm2,%xmm0 ## multiply by 4 pslld $2,%xmm2 ## move to integer registers movd %xmm2,%r12d pshufd $1,%xmm2,%xmm2 movd %xmm2,%r13d ## calculate eps subps %xmm0,%xmm3 movq nb310_VFtab(%rbp),%rsi ## load table data movlps (%rsi,%r12,4),%xmm4 movlps (%rsi,%r13,4),%xmm5 unpcklps %xmm5,%xmm4 movhlps %xmm4,%xmm5 movaps %xmm1,%xmm0 ## rinv mulps %xmm0,%xmm0 ## rinvsq movaps %xmm0,%xmm2 ## rinvsq mulps %xmm2,%xmm2 ## rinv4 mulps %xmm0,%xmm2 ## rinv6 movaps %xmm2,%xmm12 mulps %xmm12,%xmm12 ## rinv12 movlps 8(%rsi,%r12,4),%xmm6 movlps 8(%rsi,%r13,4),%xmm7 unpcklps %xmm7,%xmm6 movhlps %xmm6,%xmm7 ## table data ready in xmm4-xmm7 mulps nb310_c6(%rsp),%xmm2 ## vvdw6=c6*rinv6 mulps nb310_c12(%rsp),%xmm12 ## vvdw12=c12*rinv12 movaps %xmm12,%xmm0 subps %xmm2,%xmm12 ## Vvdw=Vvdw12-Vvdw6 ## add potential to vvdwtot addps nb310_Vvdwtot(%rsp),%xmm12 movlps %xmm12,nb310_Vvdwtot(%rsp) mulps %xmm3,%xmm7 ## Heps mulps %xmm3,%xmm6 ## Geps mulps %xmm3,%xmm7 ## Heps2 addps %xmm6,%xmm5 ## F+Geps addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm7,%xmm7 ## 2*Heps2 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps mulps %xmm3,%xmm5 ## eps*Fp addps %xmm4,%xmm5 ## VV mulps nb310_qq(%rsp),%xmm5 ## VV*qq=vcoul mulps nb310_qq(%rsp),%xmm7 ## FF*qq=fijC ## LJ forces mulps nb310_six(%rsp),%xmm2 mulps nb310_twelve(%rsp),%xmm0 subps %xmm2,%xmm0 mulps %xmm1,%xmm0 ## (12*vnb12-6*vnb6)*rinv ## add potential to vctot addps nb310_vctot(%rsp),%xmm5 movlps %xmm5,nb310_vctot(%rsp) xorps %xmm8,%xmm8 mulps nb310_tsc(%rsp),%xmm7 subps %xmm7,%xmm0 mulps %xmm1,%xmm0 ## fscal ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm0,%xmm9 mulps %xmm0,%xmm10 mulps %xmm0,%xmm11 movlhps %xmm8,%xmm9 movlhps %xmm8,%xmm10 movlhps %xmm8,%xmm11 ## accumulate i forces addps %xmm9,%xmm13 addps %xmm10,%xmm14 addps %xmm11,%xmm15 movq nb310_faction(%rbp),%rsi ## the fj's - start by accumulating x & y forces from memory movlps (%rsi,%rax,4),%xmm0 ## x1 y1 - - movhps (%rsi,%rbx,4),%xmm0 ## x1 y1 x2 y2 unpcklps %xmm10,%xmm9 ## x1 y1 x2 y2 addps %xmm9,%xmm0 movlps %xmm0,(%rsi,%rax,4) movhps %xmm0,(%rsi,%rbx,4) ## z forces pshufd $1,%xmm11,%xmm8 addss 8(%rsi,%rax,4),%xmm11 addss 8(%rsi,%rbx,4),%xmm8 movss %xmm11,8(%rsi,%rax,4) movss %xmm8,8(%rsi,%rbx,4)_nb_kernel310_x86_64_sse.nb310_checksingle: movl nb310_innerk(%rsp),%edx andl $1,%edx jnz _nb_kernel310_x86_64_sse.nb310_dosingle jmp _nb_kernel310_x86_64_sse.nb310_updateouterdata_nb_kernel310_x86_64_sse.nb310_dosingle: movq nb310_innerjjnr(%rsp),%rcx movl (%rcx),%eax movq nb310_charge(%rbp),%rsi movss (%rsi,%rax,4),%xmm0 mulss nb310_iq(%rsp),%xmm0 movaps %xmm0,nb310_qq(%rsp) movq nb310_type(%rbp),%rsi ## vdw parameters movl (%rsi,%rax,4),%r12d shll %r12d movl nb310_ntia(%rsp),%edi addl %edi,%r12d movq nb310_vdwparam(%rbp),%rsi movss (%rsi,%r12,4),%xmm0 movss 4(%rsi,%r12,4),%xmm3 movaps %xmm0,nb310_c6(%rsp) movaps %xmm3,nb310_c12(%rsp) lea (%rax,%rax,2),%rax ## replace jnr with j3 movq nb310_pos(%rbp),%rdi movss (%rdi,%rax,4),%xmm4 ## x1 - - - movss 4(%rdi,%rax,4),%xmm5 ## y2 - - - movss 8(%rdi,%rax,4),%xmm6 ## 13 - - - ## calc dr subss nb310_ix(%rsp),%xmm4 subss nb310_iy(%rsp),%xmm5 subss nb310_iz(%rsp),%xmm6 ## store dr in xmm9-xmm11 movaps %xmm4,%xmm9 movaps %xmm5,%xmm10 movaps %xmm6,%xmm11 ## square it mulss %xmm4,%xmm4 mulss %xmm5,%xmm5 mulss %xmm6,%xmm6 addss %xmm5,%xmm4 addss %xmm6,%xmm4 ## rsq in xmm4 ## calculate rinv=1/sqrt(rsq) rsqrtss %xmm4,%xmm5 movaps %xmm5,%xmm2 mulss %xmm5,%xmm5 movaps nb310_three(%rsp),%xmm1 mulss %xmm4,%xmm5 ## rsq*lu*lu subss %xmm5,%xmm1 ## 30-rsq*lu*lu mulss %xmm2,%xmm1 mulss nb310_half(%rsp),%xmm1 ## xmm1=rinv movaps %xmm4,%xmm3 ## xmm3=rsq mulss %xmm1,%xmm3 ## r mulss nb310_tsc(%rsp),%xmm3 ## rtab ## truncate and convert to integers cvttss2si %xmm3,%r12d ## convert back to float cvtsi2ss %r12d,%xmm0 ## multiply by 4 shll $2,%r12d ## calculate eps subss %xmm0,%xmm3 movq nb310_VFtab(%rbp),%rsi movaps %xmm1,%xmm0 ## rinv mulss %xmm0,%xmm0 ## rinvsq movaps %xmm0,%xmm2 ## rinvsq mulss %xmm2,%xmm2 ## rinv4 mulss %xmm0,%xmm2 ## rinv6 movaps %xmm2,%xmm12 mulss %xmm12,%xmm12 ## rinv12 ## load table data movss (%rsi,%r12,4),%xmm4 movss 4(%rsi,%r12,4),%xmm5 movss 8(%rsi,%r12,4),%xmm6 movss 12(%rsi,%r12,4),%xmm7 ## table data ready in xmm4-xmm7 mulss nb310_c6(%rsp),%xmm2 ## vvdw6=c6*rinv6 mulss nb310_c12(%rsp),%xmm12 ## vvdw12=c12*rinv12 movaps %xmm12,%xmm0 subss %xmm2,%xmm12 ## Vvdw=Vvdw12-Vvdw6 ## add potential to vvdwtot addss nb310_Vvdwtot(%rsp),%xmm12 movss %xmm12,nb310_Vvdwtot(%rsp) mulss %xmm3,%xmm7 ## Heps mulss %xmm3,%xmm6 ## Geps mulss %xmm3,%xmm7 ## Heps2 addss %xmm6,%xmm5 ## F+Geps addss %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addss %xmm7,%xmm7 ## 2*Heps2 addss %xmm6,%xmm7 ## 2*Heps2+Geps addss %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps mulss %xmm3,%xmm5 ## eps*Fp addss %xmm4,%xmm5 ## VV mulss nb310_qq(%rsp),%xmm5 ## VV*qq=vcoul mulss nb310_qq(%rsp),%xmm7 ## FF*qq=fijC ## LJ forces mulss nb310_six(%rsp),%xmm2 mulss nb310_twelve(%rsp),%xmm0 subss %xmm2,%xmm0 mulss %xmm1,%xmm0 ## (12*vnb12-6*vnb6)*rinv ## add potential to vctot addss nb310_vctot(%rsp),%xmm5 movss %xmm5,nb310_vctot(%rsp) mulss nb310_tsc(%rsp),%xmm7 subss %xmm7,%xmm0 mulss %xmm1,%xmm0 ## fscal ## calculate scalar force by multiplying dx/dy/dz with fscal mulss %xmm0,%xmm9 mulss %xmm0,%xmm10 mulss %xmm0,%xmm11 ## accumulate i forces addss %xmm9,%xmm13 addss %xmm10,%xmm14 addss %xmm11,%xmm15 movq nb310_faction(%rbp),%rsi ## add to j forces addss (%rsi,%rax,4),%xmm9 addss 4(%rsi,%rax,4),%xmm10 addss 8(%rsi,%rax,4),%xmm11 movss %xmm9,(%rsi,%rax,4) movss %xmm10,4(%rsi,%rax,4) movss %xmm11,8(%rsi,%rax,4)_nb_kernel310_x86_64_sse.nb310_updateouterdata: movl nb310_ii3(%rsp),%ecx movq nb310_faction(%rbp),%rdi movq nb310_fshift(%rbp),%rsi movl nb310_is3(%rsp),%edx ## accumulate i forces in xmm13, xmm14, xmm15 movhlps %xmm13,%xmm0 movhlps %xmm14,%xmm1⌨️ 快捷键说明
复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?