nb_kernel200_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,212 行 · 第 1/3 页
S
1,212 行
#### $Id: nb_kernel200_x86_64_sse.s,v 1.4.2.3 2006/09/22 08:32:49 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_kernel200_x86_64_sse.globl _nb_kernel200_x86_64_ssenb_kernel200_x86_64_sse: _nb_kernel200_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb200_fshift, 16.set nb200_gid, 24.set nb200_pos, 32.set nb200_faction, 40.set nb200_charge, 48.set nb200_p_facel, 56.set nb200_argkrf, 64.set nb200_argcrf, 72.set nb200_Vc, 80.set nb200_type, 88.set nb200_p_ntype, 96.set nb200_vdwparam, 104.set nb200_Vvdw, 112.set nb200_p_tabscale, 120.set nb200_VFtab, 128.set nb200_invsqrta, 136.set nb200_dvda, 144.set nb200_p_gbtabscale, 152.set nb200_GBtab, 160.set nb200_p_nthreads, 168.set nb200_count, 176.set nb200_mtx, 184.set nb200_outeriter, 192.set nb200_inneriter, 200.set nb208_work, 200 ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb200_ix, 0.set nb200_iy, 16.set nb200_iz, 32.set nb200_iq, 48.set nb200_dx, 64.set nb200_dy, 80.set nb200_dz, 96.set nb200_vctot, 112.set nb200_fix, 128.set nb200_fiy, 144.set nb200_fiz, 160.set nb200_half, 176.set nb200_three, 192.set nb200_two, 208.set nb200_krf, 224.set nb200_crf, 240.set nb200_innerjjnr, 256.set nb200_nri, 264.set nb200_iinr, 272.set nb200_jindex, 280.set nb200_jjnr, 288.set nb200_shift, 296.set nb200_shiftvec, 304.set nb200_facel, 312.set nb200_is3, 320.set nb200_ii3, 324.set nb200_innerk, 328.set nb200_n, 332.set nb200_nn1, 336.set nb200_nouter, 340.set nb200_ninner, 344 push %rbp movq %rsp,%rbp push %rbx emms push %r12 push %r13 push %r14 push %r15 subq $360,%rsp ## local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb200_nouter(%rsp) movl %eax,nb200_ninner(%rsp) movl (%rdi),%edi movl %edi,nb200_nri(%rsp) movq %rsi,nb200_iinr(%rsp) movq %rdx,nb200_jindex(%rsp) movq %rcx,nb200_jjnr(%rsp) movq %r8,nb200_shift(%rsp) movq %r9,nb200_shiftvec(%rsp) movq nb200_p_facel(%rbp),%rsi movss (%rsi),%xmm0 movss %xmm0,nb200_facel(%rsp) movq nb200_argkrf(%rbp),%rsi movq nb200_argcrf(%rbp),%rdi movss (%rsi),%xmm1 movss (%rdi),%xmm2 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm1,nb200_krf(%rsp) movaps %xmm2,nb200_crf(%rsp) ## create constant floating-point factors on stack movl $0x3f000000,%eax ## half in IEEE (hex) movl %eax,nb200_half(%rsp) movss nb200_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,nb200_half(%rsp) movaps %xmm2,nb200_two(%rsp) movaps %xmm3,nb200_three(%rsp)_nb_kernel200_x86_64_sse.nb200_threadloop: movq nb200_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel200_x86_64_sse.nb200_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_kernel200_x86_64_sse.nb200_spinlock ## if(nn1>nri) nn1=nri movl nb200_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,nb200_n(%rsp) movl %ebx,nb200_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel200_x86_64_sse.nb200_outerstart jmp _nb_kernel200_x86_64_sse.nb200_end ## assume we have at least one i particle - start directly _nb_kernel200_x86_64_sse.nb200_outerstart: ## ebx contains number of outer iterations addl nb200_nouter(%rsp),%ebx movl %ebx,nb200_nouter(%rsp)_nb_kernel200_x86_64_sse.nb200_outer: movq nb200_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,nb200_is3(%rsp) ## store is3 movq nb200_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 nb200_iinr(%rsp),%rcx ## rcx = pointer into iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb200_charge(%rbp),%rdx movss (%rdx,%rbx,4),%xmm3 mulss nb200_facel(%rsp),%xmm3 shufps $0,%xmm3,%xmm3 lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb200_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,nb200_iq(%rsp) shufps $0,%xmm0,%xmm0 shufps $0,%xmm1,%xmm1 shufps $0,%xmm2,%xmm2 movaps %xmm0,nb200_ix(%rsp) movaps %xmm1,nb200_iy(%rsp) movaps %xmm2,nb200_iz(%rsp) movl %ebx,nb200_ii3(%rsp) ## clear vctot (xmm12) and i forces (xmm13-xmm15) xorps %xmm12,%xmm12 movaps %xmm12,%xmm13 movaps %xmm12,%xmm14 movaps %xmm12,%xmm15 movq nb200_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 nb200_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb200_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb200_ninner(%rsp),%ecx movl %ecx,nb200_ninner(%rsp) addl $0,%edx movl %edx,nb200_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel200_x86_64_sse.nb200_unroll_loop jmp _nb_kernel200_x86_64_sse.nb200_finish_inner_nb_kernel200_x86_64_sse.nb200_unroll_loop: ## quad-unroll innerloop here movq nb200_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 addq $16,nb200_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%r8,%r8,2),%rax ## j3 lea (%r9,%r9,2),%rbx lea (%r10,%r10,2),%rcx lea (%r11,%r11,2),%rdx movq nb200_pos(%rbp),%rdi ## 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 movq nb200_charge(%rbp),%rsi ## calc dr subps nb200_ix(%rsp),%xmm1 subps nb200_iy(%rsp),%xmm3 subps nb200_iz(%rsp),%xmm5 ## store dr in xmm9-xmm11 movaps %xmm1,%xmm9 movaps %xmm3,%xmm10 movaps %xmm5,%xmm11 movss (%rsi,%r8,4),%xmm0 movss (%rsi,%r10,4),%xmm2 movss (%rsi,%r9,4),%xmm6 movss (%rsi,%r11,4),%xmm8 ## square it mulps %xmm1,%xmm1 mulps %xmm3,%xmm3 mulps %xmm5,%xmm5 addps %xmm3,%xmm1 addps %xmm5,%xmm1 ## rsq in xmm1 movaps nb200_krf(%rsp),%xmm7 unpcklps %xmm2,%xmm0 ## jqa jqc - - unpcklps %xmm8,%xmm6 ## jqb jqd - - rsqrtps %xmm1,%xmm5 ## lookup seed in xmm5 movaps %xmm5,%xmm2 mulps %xmm5,%xmm5 movaps nb200_three(%rsp),%xmm4 mulps %xmm1,%xmm5 ## rsq*lu*lu movaps nb200_half(%rsp),%xmm3 mulps %xmm1,%xmm7 ## xmm7=krsq subps %xmm5,%xmm4 ## 30-rsq*lu*lu mulps %xmm2,%xmm4 unpcklps %xmm6,%xmm0 ## jqa jqb jqc jqd mulps nb200_iq(%rsp),%xmm0 ##qq mulps %xmm4,%xmm3 ## xmm3=rinv movaps %xmm3,%xmm1 mulps %xmm1,%xmm1 ## xmm1=rinvsq movaps %xmm3,%xmm6 addps %xmm7,%xmm6 ## xmm6=rinv+ krsq subps nb200_crf(%rsp),%xmm6 ## xmm6=rinv+ krsq-crf mulps %xmm0,%xmm6 ## xmm6=vcoul=qq*(rinv+ krsq) addps %xmm7,%xmm7 subps %xmm7,%xmm3 mulps %xmm3,%xmm0 mulps %xmm0,%xmm1 ## xmm1=total fscal ## add potential to vctot (sum in xmm12) addps %xmm6,%xmm12 ## calculate scalar force by multiplying dx/dy/dz with fscal mulps %xmm1,%xmm9 mulps %xmm1,%xmm10 mulps %xmm1,%xmm11 movq nb200_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,nb200_innerk(%rsp) jl _nb_kernel200_x86_64_sse.nb200_finish_inner jmp _nb_kernel200_x86_64_sse.nb200_unroll_loop_nb_kernel200_x86_64_sse.nb200_finish_inner: ## check if at least two particles remain addl $4,nb200_innerk(%rsp) movl nb200_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel200_x86_64_sse.nb200_dopair jmp _nb_kernel200_x86_64_sse.nb200_checksingle_nb_kernel200_x86_64_sse.nb200_dopair: ## twice-unrolled innerloop here movq nb200_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb200_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb200_charge(%rbp),%rsi movss (%rsi,%rax,4),%xmm0⌨️ 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?