nb_kernel300_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,442 行 · 第 1/3 页
S
1,442 行
#### $Id: nb_kernel300_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:50 lindahl Exp $#### Gromacs 4.0 Copyright (c) 1991-2003 ## David van der Spoel, Erik Lindahl#### This program is free software; you can redistribute it and/or## modify it under the terms of the GNU General Public License## as published by the Free Software Foundation; either version 2## of the License, or (at your option) any later version.#### To help us fund GROMACS development, we humbly ask that you cite## the research papers on the package. Check out http://www.gromacs.org## ## And Hey:## Gnomes, ROck Monsters And Chili Sauce##.globl nb_kernel300_x86_64_sse.globl _nb_kernel300_x86_64_ssenb_kernel300_x86_64_sse: _nb_kernel300_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb300_fshift, 16.set nb300_gid, 24.set nb300_pos, 32.set nb300_faction, 40.set nb300_charge, 48.set nb300_p_facel, 56.set nb300_argkrf, 64.set nb300_argcrf, 72.set nb300_Vc, 80.set nb300_type, 88.set nb300_p_ntype, 96.set nb300_vdwparam, 104.set nb300_Vvdw, 112.set nb300_p_tabscale, 120.set nb300_VFtab, 128.set nb300_invsqrta, 136.set nb300_dvda, 144.set nb300_p_gbtabscale, 152.set nb300_GBtab, 160.set nb300_p_nthreads, 168.set nb300_count, 176.set nb300_mtx, 184.set nb300_outeriter, 192.set nb300_inneriter, 200.set nb300_work, 208 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb300_ix, 0.set nb300_iy, 16.set nb300_iz, 32.set nb300_iq, 48.set nb300_dx, 64.set nb300_dy, 80.set nb300_dz, 96.set nb300_two, 112.set nb300_tsc, 128.set nb300_qq, 144.set nb300_fs, 160.set nb300_vctot, 176.set nb300_fix, 192.set nb300_fiy, 208.set nb300_fiz, 224.set nb300_half, 240.set nb300_three, 256.set nb300_innerjjnr, 272.set nb300_nri, 280.set nb300_iinr, 288.set nb300_jindex, 296.set nb300_jjnr, 304.set nb300_shift, 312.set nb300_shiftvec, 320.set nb300_facel, 328.set nb300_innerk, 336.set nb300_is3, 344.set nb300_ii3, 348.set nb300_n, 352.set nb300_nn1, 356.set nb300_nouter, 360.set nb300_ninner, 364 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $376,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb300_nouter(%rsp) movl %eax,nb300_ninner(%rsp) movl (%rdi),%edi movl %edi,nb300_nri(%rsp) movq %rsi,nb300_iinr(%rsp) movq %rdx,nb300_jindex(%rsp) movq %rcx,nb300_jjnr(%rsp) movq %r8,nb300_shift(%rsp) movq %r9,nb300_shiftvec(%rsp) movq nb300_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb300_facel(%rsp) movq nb300_p_tabscale(%rbp),%rax movss (%rax),%xmm3 shufps $0,%xmm3,%xmm3 movaps %xmm3,nb300_tsc(%rsp) movq nb300_pos(%rbp),%r8 movq nb300_faction(%rbp),%r9 movq nb300_charge(%rbp),%r10 movq nb300_VFtab(%rbp),%r11 ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb300_half(%rsp) movss nb300_half(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm2,%xmm2 ## one movaps %xmm2,%xmm3 addps %xmm2,%xmm2 ## two addps %xmm2,%xmm3 ## three movaps %xmm1,nb300_half(%rsp) movaps %xmm2,nb300_two(%rsp) movaps %xmm3,nb300_three(%rsp)_nb_kernel300_x86_64_sse.nb300_threadloop: movq nb300_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel300_x86_64_sse.nb300_spinlock: movl %eax,%ebx ## ebx=*count=nn0 addl $1,%ebx ## ebx=nn1=nn0+10 lock cmpxchgl %ebx,(%rsi) ## write nn1 to *counter, ## if it hasnt changed. ## or reread *counter to eax. pause ## -> better p4 performance jnz _nb_kernel300_x86_64_sse.nb300_spinlock ## if(nn1>nri) nn1=nri movl nb300_nri(%rsp),%ecx movl %ecx,%edx subl %ebx,%ecx cmovlel %edx,%ebx ## if(nn1>nri) nn1=nri ## Cleared the spinlock if we got here. ## eax contains nn0, ebx contains nn1. movl %eax,nb300_n(%rsp) movl %ebx,nb300_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel300_x86_64_sse.nb300_outerstart jmp _nb_kernel300_x86_64_sse.nb300_end_nb_kernel300_x86_64_sse.nb300_outerstart: ## ebx contains number of outer iterations addl nb300_nouter(%rsp),%ebx movl %ebx,nb300_nouter(%rsp)_nb_kernel300_x86_64_sse.nb300_outer: movq nb300_shift(%rsp),%rax ## rax = pointer into shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb300_is3(%rsp) ## store is3 movq nb300_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm0 movss 4(%rax,%rbx,4),%xmm1 movss 8(%rax,%rbx,4),%xmm2 movq nb300_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb300_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb300_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb300_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm0 addss 4(%rax,%rbx,4),%xmm1 addss 8(%rax,%rbx,4),%xmm2 movaps %xmm3,nb300_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb300_ix(%rsp) movaps %xmm1,nb300_iy(%rsp) movaps %xmm2,nb300_iz(%rsp) movl %ebx,nb300_ii3(%rsp) ## clear vctot (xmm12) and i forces (xmm13-xmm15) xorps %xmm12,%xmm12 movaps %xmm12,%xmm13 movaps %xmm12,%xmm14 movaps %xmm12,%xmm15 movq nb300_jindex(%rsp),%rax movl (%rax,%rsi,4),%ecx ## jindex[n] movl 4(%rax,%rsi,4),%edx ## jindex[n+1] subl %ecx,%edx ## number of innerloop atoms movq nb300_pos(%rbp),%rsi movq nb300_faction(%rbp),%rdi movq nb300_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb300_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb300_ninner(%rsp),%ecx movl %ecx,nb300_ninner(%rsp) addl $0,%edx movl %edx,nb300_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel300_x86_64_sse.nb300_unroll_loop jmp _nb_kernel300_x86_64_sse.nb300_finish_inner_nb_kernel300_x86_64_sse.nb300_unroll_loop: ## quad-unrolled innerloop here movq nb300_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%r8d movl 4(%rdx),%r9d movl 8(%rdx),%r10d movl 12(%rdx),%r11d ## eax-edx=jnr1-4 movq nb300_pos(%rbp),%rdi addq $16,nb300_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%r8,%r8,2),%rax ## replace jnr with j3 lea (%r9,%r9,2),%rbx lea (%r10,%r10,2),%rcx ## replace jnr with j3 lea (%r11,%r11,2),%rdx ## load coordinates movlps (%rdi,%rax,4),%xmm1 ## x1 y1 - - movlps (%rdi,%rbx,4),%xmm2 ## x2 y2 - - movlps (%rdi,%rcx,4),%xmm3 ## x3 y3 - - movlps (%rdi,%rdx,4),%xmm4 ## x4 y4 - - movss 8(%rdi,%rax,4),%xmm5 ## z1 - - - movss 8(%rdi,%rbx,4),%xmm6 ## z2 - - - movss 8(%rdi,%rcx,4),%xmm7 ## z3 - - - movss 8(%rdi,%rdx,4),%xmm8 ## z4 - - - unpcklps %xmm3,%xmm1 ## x1 x3 y1 y3 unpcklps %xmm4,%xmm2 ## x2 x4 y2 y4 unpcklps %xmm7,%xmm5 ## z1 z3 - - unpcklps %xmm8,%xmm6 ## z2 z4 - - movaps %xmm1,%xmm3 unpcklps %xmm2,%xmm1 ## x1 x2 x3 x4 unpckhps %xmm2,%xmm3 ## y1 y2 y3 y4 unpcklps %xmm6,%xmm5 ## z1 z2 z3 z4 ## calc dr subps nb300_ix(%rsp),%xmm1 subps nb300_iy(%rsp),%xmm3 subps nb300_iz(%rsp),%xmm5 ## store dr in xmm9-xmm11 movaps %xmm1,%xmm9 movaps %xmm3,%xmm10 movaps %xmm5,%xmm11 ## square it mulps %xmm1,%xmm1 mulps %xmm3,%xmm3 mulps %xmm5,%xmm5 addps %xmm1,%xmm3 addps %xmm5,%xmm3 ## rsq in xmm3 ## calculate rinv=1/sqrt(rsq) rsqrtps %xmm3,%xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb300_three(%rsp),%xmm1 mulps %xmm3,%xmm5 ## rsq*lu*lu subps %xmm5,%xmm1 ## 30-rsq*lu*lu mulps %xmm2,%xmm1 mulps nb300_half(%rsp),%xmm1 ## xmm1=rinv ## xmm3=rsq mulps %xmm1,%xmm3 ## r mulps nb300_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 movhlps %xmm2,%xmm7 movd %xmm2,%r12d movd %xmm7,%r14d pshufd $1,%xmm2,%xmm2 pshufd $1,%xmm7,%xmm7 movd %xmm2,%r13d movd %xmm7,%r15d ## calculate eps subps %xmm0,%xmm3 movq nb300_VFtab(%rbp),%rsi ## load table data movlps (%rsi,%r12,4),%xmm5 movlps (%rsi,%r14,4),%xmm7 movhps (%rsi,%r13,4),%xmm5 movhps (%rsi,%r15,4),%xmm7 movaps %xmm5,%xmm4 shufps $136,%xmm7,%xmm4 ## 10001000 shufps $221,%xmm7,%xmm5 ## 11011101 movlps 8(%rsi,%r12,4),%xmm7 movlps 8(%rsi,%r14,4),%xmm8 movhps 8(%rsi,%r13,4),%xmm7 movhps 8(%rsi,%r15,4),%xmm8 movaps %xmm7,%xmm6 movq nb300_charge(%rbp),%rdi shufps $136,%xmm8,%xmm6 ## 10001000 shufps $221,%xmm8,%xmm7 ## 11011101 ## table data ready in xmm4-xmm7 movss (%rdi,%r8,4),%xmm0 movss (%rdi,%r10,4),%xmm2 mulps %xmm3,%xmm7 ## Heps mulps %xmm3,%xmm6 ## Geps mulps %xmm3,%xmm7 ## Heps2 unpcklps %xmm2,%xmm0 movss (%rdi,%r9,4),%xmm2 movss (%rdi,%r11,4),%xmm8 addps %xmm6,%xmm5 ## F+Geps addps %xmm7,%xmm5 ## F+Geps+Heps2 = Fp addps %xmm7,%xmm7 ## 2*Heps2 unpcklps %xmm8,%xmm2 unpcklps %xmm2,%xmm0 addps %xmm6,%xmm7 ## 2*Heps2+Geps addps %xmm5,%xmm7 ## FF = Fp + 2*Heps2 + Geps mulps %xmm3,%xmm5 ## eps*Fp mulps nb300_iq(%rsp),%xmm0 addps %xmm4,%xmm5 ## VV mulps %xmm0,%xmm5 ## VV*qq=vcoul mulps %xmm0,%xmm7 ## FF*qq=fijC ## add potential to vctot (sum in xmm12) addps %xmm5,%xmm12 mulps nb300_tsc(%rsp),%xmm7 mulps %xmm1,%xmm7 xorps %xmm4,%xmm4 subps %xmm7,%xmm4 ## fscal movq nb300_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 ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm4,%xmm9 mulps %xmm4,%xmm10 mulps %xmm4,%xmm11 ## 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,nb300_innerk(%rsp) jl _nb_kernel300_x86_64_sse.nb300_finish_inner jmp _nb_kernel300_x86_64_sse.nb300_unroll_loop_nb_kernel300_x86_64_sse.nb300_finish_inner: ## check if at least two particles remain addl $4,nb300_innerk(%rsp) movl nb300_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel300_x86_64_sse.nb300_dopair jmp _nb_kernel300_x86_64_sse.nb300_checksingle_nb_kernel300_x86_64_sse.nb300_dopair: ## twice-unrolled innerloop here movq nb300_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb300_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb300_charge(%rbp),%rsi movss (%rsi,%rax,4),%xmm8 movss (%rsi,%rbx,4),%xmm1 unpcklps %xmm1,%xmm8 ## jqa jqb - - mulps nb300_iq(%rsp),%xmm8 ##qq lea (%rax,%rax,2),%rax ## replace jnr with j3 lea (%rbx,%rbx,2),%rbx ## load coordinates movq nb300_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 nb300_ix(%rsp),%xmm4 subps nb300_iy(%rsp),%xmm5 subps nb300_iz(%rsp),%xmm6 ## store dr in xmm9-xmm11⌨️ 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?