nb_kernel111_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,633 行 · 第 1/4 页
S
1,633 行
movaps %xmm7,nb111_qqH(%rsp) movq nb111_type(%rbp),%rsi movl (%rsi,%r8,4),%r8d movl (%rsi,%r9,4),%r9d movl (%rsi,%r10,4),%r10d movl (%rsi,%r11,4),%r11d ## start doing invsqrt rsqrtps %xmm0,%xmm1 rsqrtps %xmm3,%xmm4 rsqrtps %xmm6,%xmm7 movaps %xmm1,%xmm2 movaps %xmm4,%xmm5 movaps %xmm7,%xmm8 mulps %xmm1,%xmm1 ## lu*lu mulps %xmm4,%xmm4 ## lu*lu mulps %xmm7,%xmm7 ## lu*lu movq nb111_vdwparam(%rbp),%rsi shll %r8d shll %r9d shll %r10d shll %r11d movaps nb111_three(%rsp),%xmm9 movaps %xmm9,%xmm10 movaps %xmm9,%xmm11 mulps %xmm0,%xmm1 ## rsq*lu*lu mulps %xmm3,%xmm4 ## rsq*lu*lu mulps %xmm6,%xmm7 ## rsq*lu*lu subps %xmm1,%xmm9 subps %xmm4,%xmm10 subps %xmm7,%xmm11 ## 3-rsq*lu*lu movl nb111_ntia(%rsp),%edi addl %edi,%r8d addl %edi,%r9d addl %edi,%r10d addl %edi,%r11d mulps %xmm2,%xmm9 mulps %xmm5,%xmm10 mulps %xmm8,%xmm11 ## lu*(3-rsq*lu*lu) movaps nb111_half(%rsp),%xmm0 mulps %xmm0,%xmm9 ## rinvO mulps %xmm0,%xmm10 ## rinvH1 mulps %xmm0,%xmm11 ## rinvH2 movlps (%rsi,%r8,4),%xmm3 movlps (%rsi,%r10,4),%xmm2 movhps (%rsi,%r9,4),%xmm3 movhps (%rsi,%r11,4),%xmm2 movaps %xmm3,%xmm5 shufps $136,%xmm2,%xmm3 ## 10001000 shufps $221,%xmm2,%xmm5 ## 11011101 ## interactions movaps %xmm9,%xmm0 movaps %xmm10,%xmm1 movaps %xmm11,%xmm2 mulps %xmm9,%xmm9 ## rinvsq mulps %xmm10,%xmm10 mulps %xmm11,%xmm11 movaps %xmm9,%xmm12 mulps %xmm12,%xmm12 ## rinv4 mulps %xmm9,%xmm12 ## rinv6 mulps nb111_qqO(%rsp),%xmm0 mulps nb111_qqH(%rsp),%xmm1 mulps nb111_qqH(%rsp),%xmm2 movaps %xmm12,%xmm13 ## rinv6 mulps %xmm12,%xmm12 ## rinv12 mulps %xmm3,%xmm13 mulps %xmm5,%xmm12 movaps %xmm12,%xmm14 subps %xmm13,%xmm14 addps nb111_Vvdwtot(%rsp),%xmm14 mulps nb111_six(%rsp),%xmm13 mulps nb111_twelve(%rsp),%xmm12 movaps %xmm14,nb111_Vvdwtot(%rsp) subps %xmm13,%xmm12 ## LJ fscal addps %xmm0,%xmm12 mulps %xmm12,%xmm9 mulps %xmm1,%xmm10 mulps %xmm2,%xmm11 addps nb111_vctot(%rsp),%xmm0 addps %xmm2,%xmm1 addps %xmm1,%xmm0 movaps %xmm0,nb111_vctot(%rsp) movq nb111_faction(%rbp),%rdi ## move j forces to local temp variables movlps (%rdi,%rax,4),%xmm0 ## jxa jya - - movlps (%rdi,%rcx,4),%xmm1 ## jxc jyc - - movhps (%rdi,%rbx,4),%xmm0 ## jxa jya jxb jyb movhps (%rdi,%rdx,4),%xmm1 ## jxc jyc jxd jyd movss 8(%rdi,%rax,4),%xmm2 ## jza - - - movss 8(%rdi,%rcx,4),%xmm3 ## jzc - - - movss 8(%rdi,%rbx,4),%xmm4 ## jzb movss 8(%rdi,%rdx,4),%xmm5 ## jzd movlhps %xmm4,%xmm2 ## jza - jzb - movlhps %xmm5,%xmm3 ## jzc - jzd - shufps $136,%xmm3,%xmm2 ## 10001000 => jza jzb jzc jzd ## xmm0: jxa jya jxb jyb ## xmm1: jxc jyc jxd jyd ## xmm2: jza jzb jzc jzd movaps %xmm9,%xmm7 movaps %xmm9,%xmm8 movaps %xmm11,%xmm13 movaps %xmm11,%xmm14 movaps %xmm11,%xmm15 movaps %xmm10,%xmm11 movaps %xmm10,%xmm12 mulps nb111_dxO(%rsp),%xmm7 mulps nb111_dyO(%rsp),%xmm8 mulps nb111_dzO(%rsp),%xmm9 mulps nb111_dxH1(%rsp),%xmm10 mulps nb111_dyH1(%rsp),%xmm11 mulps nb111_dzH1(%rsp),%xmm12 mulps nb111_dxH2(%rsp),%xmm13 mulps nb111_dyH2(%rsp),%xmm14 mulps nb111_dzH2(%rsp),%xmm15 movaps %xmm7,%xmm3 movaps %xmm8,%xmm4 addps %xmm9,%xmm2 addps nb111_fixO(%rsp),%xmm7 addps nb111_fiyO(%rsp),%xmm8 addps nb111_fizO(%rsp),%xmm9 addps %xmm10,%xmm3 addps %xmm11,%xmm4 addps %xmm12,%xmm2 addps nb111_fixH1(%rsp),%xmm10 addps nb111_fiyH1(%rsp),%xmm11 addps nb111_fizH1(%rsp),%xmm12 addps %xmm13,%xmm3 addps %xmm14,%xmm4 addps %xmm15,%xmm2 addps nb111_fixH2(%rsp),%xmm13 addps nb111_fiyH2(%rsp),%xmm14 addps nb111_fizH2(%rsp),%xmm15 movaps %xmm7,nb111_fixO(%rsp) movaps %xmm8,nb111_fiyO(%rsp) movaps %xmm9,nb111_fizO(%rsp) movaps %xmm10,nb111_fixH1(%rsp) movaps %xmm11,nb111_fiyH1(%rsp) movaps %xmm12,nb111_fizH1(%rsp) movaps %xmm13,nb111_fixH2(%rsp) movaps %xmm14,nb111_fiyH2(%rsp) movaps %xmm15,nb111_fizH2(%rsp) ## xmm3 = fjx , xmm4 = fjy, xmm5=fjz movaps %xmm3,%xmm5 unpcklps %xmm4,%xmm3 unpckhps %xmm4,%xmm5 addps %xmm3,%xmm0 addps %xmm5,%xmm1 movhlps %xmm2,%xmm3 movlps %xmm0,(%rdi,%rax,4) movhps %xmm0,(%rdi,%rbx,4) movlps %xmm1,(%rdi,%rcx,4) movhps %xmm1,(%rdi,%rdx,4) movss %xmm2,8(%rdi,%rax,4) movss %xmm3,8(%rdi,%rcx,4) shufps $1,%xmm2,%xmm2 shufps $1,%xmm3,%xmm3 movss %xmm2,8(%rdi,%rbx,4) movss %xmm3,8(%rdi,%rdx,4) ## should we do one more iteration? subl $4,nb111_innerk(%rsp) jl _nb_kernel111_x86_64_sse.nb111_odd_inner jmp _nb_kernel111_x86_64_sse.nb111_unroll_loop_nb_kernel111_x86_64_sse.nb111_odd_inner: addl $4,nb111_innerk(%rsp) jnz _nb_kernel111_x86_64_sse.nb111_odd_loop jmp _nb_kernel111_x86_64_sse.nb111_updateouterdata_nb_kernel111_x86_64_sse.nb111_odd_loop: movq nb111_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax addq $4,nb111_innerjjnr(%rsp) xorps %xmm4,%xmm4 movss nb111_iqO(%rsp),%xmm4 movq nb111_charge(%rbp),%rsi movhps nb111_iqH(%rsp),%xmm4 movss (%rsi,%rax,4),%xmm3 ## charge in xmm3 shufps $0,%xmm3,%xmm3 mulps %xmm4,%xmm3 movaps %xmm3,nb111_qqO(%rsp) ## use oxygen qq for storage xorps %xmm6,%xmm6 movq nb111_type(%rbp),%rsi movl (%rsi,%rax,4),%ebx movq nb111_vdwparam(%rbp),%rsi shll %ebx addl nb111_ntia(%rsp),%ebx movlps (%rsi,%rbx,4),%xmm6 movaps %xmm6,%xmm7 shufps $252,%xmm6,%xmm6 ## 11111100 shufps $253,%xmm7,%xmm7 ## 11111101 movaps %xmm6,nb111_c6(%rsp) movaps %xmm7,nb111_c12(%rsp) movq nb111_pos(%rbp),%rsi lea (%rax,%rax,2),%rax ## move j coords to xmm0-xmm2 movss (%rsi,%rax,4),%xmm3 movss 4(%rsi,%rax,4),%xmm4 movss 8(%rsi,%rax,4),%xmm5 shufps $0,%xmm3,%xmm3 shufps $0,%xmm4,%xmm4 shufps $0,%xmm5,%xmm5 movss nb111_ixO(%rsp),%xmm0 movss nb111_iyO(%rsp),%xmm1 movss nb111_izO(%rsp),%xmm2 movlps nb111_ixH1(%rsp),%xmm6 movlps nb111_ixH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm0 movlps nb111_iyH1(%rsp),%xmm6 movlps nb111_iyH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm1 movlps nb111_izH1(%rsp),%xmm6 movlps nb111_izH2(%rsp),%xmm7 unpcklps %xmm7,%xmm6 movlhps %xmm6,%xmm2 subps %xmm0,%xmm3 subps %xmm1,%xmm4 subps %xmm2,%xmm5 movaps %xmm3,nb111_dxO(%rsp) movaps %xmm4,nb111_dyO(%rsp) movaps %xmm5,nb111_dzO(%rsp) mulps %xmm3,%xmm3 mulps %xmm4,%xmm4 mulps %xmm5,%xmm5 addps %xmm3,%xmm4 addps %xmm5,%xmm4 ## rsq in xmm4 rsqrtps %xmm4,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb111_three(%rsp),%xmm1 mulps %xmm4,%xmm5 ## rsq*lu*lu movaps nb111_half(%rsp),%xmm0 subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps %xmm1,%xmm0 ## xmm0=rinv ## a little trick to avoid NaNs: ## positions 0,2,and 3 are valid, but not 1. ## If it contains NaN it doesnt help to mult by 0, ## So we shuffle it and copy pos 0 to pos1! shufps $224,%xmm0,%xmm0 ## 11100000 movaps %xmm0,%xmm4 mulps %xmm4,%xmm4 ## xmm4=rinvsq movaps %xmm4,%xmm1 mulss %xmm4,%xmm1 movaps nb111_qqO(%rsp),%xmm3 mulss %xmm4,%xmm1 ## xmm1=rinvsix movaps %xmm1,%xmm2 mulss %xmm2,%xmm2 ## xmm2=rinvtwelve mulps %xmm0,%xmm3 ## xmm3=vcoul mulps nb111_c6(%rsp),%xmm1 mulps nb111_c12(%rsp),%xmm2 movaps %xmm2,%xmm5 subss %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 addps nb111_Vvdwtot(%rsp),%xmm5 mulss nb111_six(%rsp),%xmm1 mulss nb111_twelve(%rsp),%xmm2 subss %xmm1,%xmm2 addps %xmm3,%xmm2 mulps %xmm2,%xmm4 ## xmm4=total fscal addps nb111_vctot(%rsp),%xmm3 movaps nb111_dxO(%rsp),%xmm0 movaps nb111_dyO(%rsp),%xmm1 movaps nb111_dzO(%rsp),%xmm2 movaps %xmm3,nb111_vctot(%rsp) movaps %xmm5,nb111_Vvdwtot(%rsp) mulps %xmm4,%xmm0 mulps %xmm4,%xmm1 mulps %xmm4,%xmm2 ## xmm0-xmm2 contains tx-tz (partial force) movss nb111_fixO(%rsp),%xmm3 movss nb111_fiyO(%rsp),%xmm4 movss nb111_fizO(%rsp),%xmm5 addss %xmm0,%xmm3 addss %xmm1,%xmm4 addss %xmm2,%xmm5 movss %xmm3,nb111_fixO(%rsp) movss %xmm4,nb111_fiyO(%rsp) movss %xmm5,nb111_fizO(%rsp) ## updated the O force now do the H's movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $230,%xmm3,%xmm3 ## 11100110 ;# shift right shufps $230,%xmm4,%xmm4 ## 11100110 shufps $230,%xmm5,%xmm5 ## 11100110 addss nb111_fixH1(%rsp),%xmm3 addss nb111_fiyH1(%rsp),%xmm4 addss nb111_fizH1(%rsp),%xmm5 movss %xmm3,nb111_fixH1(%rsp) movss %xmm4,nb111_fiyH1(%rsp) movss %xmm5,nb111_fizH1(%rsp) ## updated the H1 force movq nb111_faction(%rbp),%rdi shufps $231,%xmm3,%xmm3 ## 11100111 ;# shift right shufps $231,%xmm4,%xmm4 ## 11100111 shufps $231,%xmm5,%xmm5 ## 11100111 addss nb111_fixH2(%rsp),%xmm3 addss nb111_fiyH2(%rsp),%xmm4 addss nb111_fizH2(%rsp),%xmm5 movss %xmm3,nb111_fixH2(%rsp) movss %xmm4,nb111_fiyH2(%rsp) movss %xmm5,nb111_fizH2(%rsp) ## updated the H2 force ## the fj's - start by accumulating the tx/ty/tz force in xmm0, xmm1 xorps %xmm5,%xmm5 movaps %xmm0,%xmm3 movlps (%rdi,%rax,4),%xmm6 movss 8(%rdi,%rax,4),%xmm7 unpcklps %xmm1,%xmm3 movlhps %xmm5,%xmm3 unpckhps %xmm1,%xmm0 addps %xmm3,%xmm0 movhlps %xmm0,%xmm3 addps %xmm3,%xmm0 ## x,y sum in xmm0 movhlps %xmm2,%xmm1 addss %xmm1,%xmm2 shufps $1,%xmm1,%xmm1 addss %xmm1,%xmm2 ## z sum in xmm2 addps %xmm0,%xmm6 addss %xmm2,%xmm7 movlps %xmm6,(%rdi,%rax,4) movss %xmm7,8(%rdi,%rax,4) decl nb111_innerk(%rsp) jz _nb_kernel111_x86_64_sse.nb111_updateouterdata jmp _nb_kernel111_x86_64_sse.nb111_odd_loop_nb_kernel111_x86_64_sse.nb111_updateouterdata: movl nb111_ii3(%rsp),%ecx movq nb111_faction(%rbp),%rdi movq nb111_fshift(%rbp),%rsi movl nb111_is3(%rsp),%edx ## accumulate Oi forces in xmm0, xmm1, xmm2 movaps nb111_fixO(%rsp),%xmm0 movaps nb111_fiyO(%rsp),%xmm1 movaps nb111_fizO(%rsp),%xmm2 movhlps %xmm0,%xmm3 movhlps %xmm1,%xmm4 movhlps %xmm2,%xmm5 addps %xmm3,%xmm0 addps %xmm4,%xmm1 addps %xmm5,%xmm2 ## sum is in 1/2 in xmm0-xmm2 movaps %xmm0,%xmm3 movaps %xmm1,%xmm4 movaps %xmm2,%xmm5 shufps $1,%xmm3,%xmm3 shufps $1,%xmm4,%xmm4 shufps $1,%xmm5,%xmm5 addss %xmm3,%xmm0 addss %xmm4,%xmm1 addss %xmm5,%xmm2 ## xmm0-xmm2 has single force in pos0 ## increment i force movss (%rdi,%rcx,4),%xmm3 movss 4(%rdi,%rcx,4),%xmm4 movss 8(%rdi,%rcx,4),%xmm5 subss %xmm0,%xmm3⌨️ 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?