nb_kernel010_x86_64_sse.s
来自「最著名最快的分子模拟软件」· S 代码 · 共 1,275 行 · 第 1/3 页
S
1,275 行
#### $Id: nb_kernel010_x86_64_sse.s,v 1.5.2.3 2006/09/22 08:32:48 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#### nb010 - forces are calculated.globl nb_kernel010_x86_64_sse.globl _nb_kernel010_x86_64_ssenb_kernel010_x86_64_sse: _nb_kernel010_x86_64_sse: ## Room for return address and rbp (16 bytes).set nb010_fshift, 16.set nb010_gid, 24.set nb010_pos, 32.set nb010_faction, 40.set nb010_charge, 48.set nb010_p_facel, 56.set nb010_argkrf, 64.set nb010_argcrf, 72.set nb010_Vc, 80.set nb010_type, 88.set nb010_p_ntype, 96.set nb010_vdwparam, 104.set nb010_Vvdw, 112.set nb010_p_tabscale, 120.set nb010_VFtab, 128.set nb010_invsqrta, 136.set nb010_dvda, 144.set nb010_p_gbtabscale, 152.set nb010_GBtab, 160.set nb010_p_nthreads, 168.set nb010_count, 176.set nb010_mtx, 184.set nb010_outeriter, 192.set nb010_inneriter, 200.set nb010_work, 208 ## The mutex (last arg) is not used in assembly. ## stack offsets for local variables ## bottom of stack is cache-aligned for sse use .set nb010_ix, 0.set nb010_iy, 16.set nb010_iz, 32.set nb010_dx, 48.set nb010_dy, 64.set nb010_dz, 80.set nb010_two, 96.set nb010_c6, 112.set nb010_c12, 128.set nb010_six, 144.set nb010_twelve, 160.set nb010_Vvdwtot, 176.set nb010_fix, 192.set nb010_fiy, 208.set nb010_fiz, 224.set nb010_half, 240.set nb010_three, 256.set nb010_nri, 272.set nb010_iinr, 280.set nb010_jindex, 288.set nb010_jjnr, 296.set nb010_shift, 304.set nb010_shiftvec, 312.set nb010_facel, 320.set nb010_innerjjnr, 328.set nb010_is3, 336.set nb010_ii3, 340.set nb010_ntia, 344.set nb010_innerk, 348.set nb010_n, 352.set nb010_nn1, 356.set nb010_ntype, 360.set nb010_nouter, 364.set nb010_ninner, 368 push %rbp movq %rsp,%rbp push %rbx push %r12 push %r13 push %r14 push %r15 emms subq $392,%rsp # # local variable stack space (n*16+8) ## zero 32-bit iteration counters movl $0,%eax movl %eax,nb010_nouter(%rsp) movl %eax,nb010_ninner(%rsp) movl (%rdi),%edi movl %edi,nb010_nri(%rsp) movq %rsi,nb010_iinr(%rsp) movq %rdx,nb010_jindex(%rsp) movq %rcx,nb010_jjnr(%rsp) movq %r8,nb010_shift(%rsp) movq %r9,nb010_shiftvec(%rsp) movq nb010_p_ntype(%rbp),%rdi movl (%rdi),%edi movl %edi,nb010_ntype(%rsp) ## create constant floating-point factors on stack movl $0x40000000,%eax ## 2.0 in IEEE (hex) movl %eax,nb010_two(%rsp) movss nb010_two(%rsp),%xmm1 shufps $0,%xmm1,%xmm1 ## splat to all elements movaps %xmm1,%xmm2 addps %xmm1,%xmm2 ## 4.0 addps %xmm1,%xmm2 ## 6.0 movaps %xmm2,%xmm3 addps %xmm3,%xmm3 ## 12.0 movaps %xmm1,nb010_two(%rsp) movaps %xmm2,nb010_six(%rsp) movaps %xmm3,nb010_twelve(%rsp)_nb_kernel010_x86_64_sse.nb010_threadloop: movq nb010_count(%rbp),%rsi ## pointer to sync counter movl (%rsi),%eax_nb_kernel010_x86_64_sse.nb010_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_kernel010_x86_64_sse.nb010_spinlock ## if(nn1>nri) nn1=nri movl nb010_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,nb010_n(%rsp) movl %ebx,nb010_nn1(%rsp) subl %eax,%ebx ## calc number of outer lists movl %eax,%esi ## copy n to esi jg _nb_kernel010_x86_64_sse.nb010_outerstart jmp _nb_kernel010_x86_64_sse.nb010_end_nb_kernel010_x86_64_sse.nb010_outerstart: ## ebx contains number of outer iterations addl nb010_nouter(%rsp),%ebx movl %ebx,nb010_nouter(%rsp)_nb_kernel010_x86_64_sse.nb010_outer: movq nb010_shift(%rsp),%rax ## rax = base of shift[] movl (%rax,%rsi,4),%ebx ## ebx=shift[n] lea (%rbx,%rbx,2),%rbx ## rbx=3*is movl %ebx,nb010_is3(%rsp) ## store is3 movq nb010_shiftvec(%rsp),%rax ## rax = base of shiftvec[] movss (%rax,%rbx,4),%xmm10 movss 4(%rax,%rbx,4),%xmm11 movss 8(%rax,%rbx,4),%xmm12 movq nb010_iinr(%rsp),%rcx ## rcx = base of iinr[] movl (%rcx,%rsi,4),%ebx ## ebx =ii movq nb010_type(%rbp),%rdx movl (%rdx,%rbx,4),%edx imull nb010_ntype(%rsp),%edx shll %edx movl %edx,nb010_ntia(%rsp) lea (%rbx,%rbx,2),%rbx ## rbx = 3*ii=ii3 movq nb010_pos(%rbp),%rax ## rax = base of pos[] addss (%rax,%rbx,4),%xmm10 addss 4(%rax,%rbx,4),%xmm11 addss 8(%rax,%rbx,4),%xmm12 shufps $0,%xmm10,%xmm10 shufps $0,%xmm11,%xmm11 shufps $0,%xmm12,%xmm12 movaps %xmm10,nb010_ix(%rsp) movaps %xmm11,nb010_iy(%rsp) movaps %xmm12,nb010_iz(%rsp) movl %ebx,nb010_ii3(%rsp) ## clear vvdwtot (xmm12) and i forces (xmm13-xmm15) xorps %xmm12,%xmm12 movaps %xmm12,%xmm13 movaps %xmm12,%xmm14 movaps %xmm12,%xmm15 movq nb010_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 nb010_jjnr(%rsp),%rax shll $2,%ecx addq %rcx,%rax movq %rax,nb010_innerjjnr(%rsp) ## pointer to jjnr[nj0] movl %edx,%ecx subl $4,%edx addl nb010_ninner(%rsp),%ecx movl %ecx,nb010_ninner(%rsp) addl $0,%edx movl %edx,nb010_innerk(%rsp) ## number of innerloop atoms jge _nb_kernel010_x86_64_sse.nb010_unroll_loop jmp _nb_kernel010_x86_64_sse.nb010_finish_inner_nb_kernel010_x86_64_sse.nb010_unroll_loop: ## quad-unrolled innerloop here movq nb010_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx movl 8(%rdx),%ecx movl 12(%rdx),%edx ## eax-edx=jnr1-4 addq $16,nb010_innerjjnr(%rsp) ## advance pointer (unrolled 4) lea (%rax,%rax,2),%r8 ## replace jnr with j3 lea (%rbx,%rbx,2),%r9 lea (%rcx,%rcx,2),%r10 lea (%rdx,%rdx,2),%r11 movq nb010_pos(%rbp),%rdi ## load coordinates movlps (%rdi,%r8,4),%xmm1 ## x1 y1 - - movlps (%rdi,%r10,4),%xmm2 ## x3 y3 - - movhps (%rdi,%r9,4),%xmm1 ## x2 y2 - - movhps (%rdi,%r11,4),%xmm2 ## x4 y4 - - movss 8(%rdi,%r8,4),%xmm5 ## z1 - - - movss 8(%rdi,%r10,4),%xmm6 ## z2 - - - movss 8(%rdi,%r9,4),%xmm7 ## z3 - - - movss 8(%rdi,%r11,4),%xmm8 ## z4 - - - movlhps %xmm7,%xmm5 ## jzOa - jzOb - movlhps %xmm8,%xmm6 ## jzOc - jzOd - movq nb010_type(%rbp),%rsi movaps %xmm1,%xmm4 unpcklps %xmm2,%xmm1 ## jxa jxc jya jyc unpckhps %xmm2,%xmm4 ## jxb jxd jyb jyd movaps %xmm1,%xmm2 unpcklps %xmm4,%xmm1 ## x unpckhps %xmm4,%xmm2 ## y shufps $136,%xmm6,%xmm5 ## 10001000 => jzH2a jzH2b jzH2c jzH2d ## load vdw types movl (%rsi,%rax,4),%r12d movl (%rsi,%rbx,4),%r13d movl (%rsi,%rcx,4),%r14d movl (%rsi,%rdx,4),%r15d ## calc dr subps nb010_ix(%rsp),%xmm1 subps nb010_iy(%rsp),%xmm2 subps nb010_iz(%rsp),%xmm5 ## store dr in xmm9-xmm11 movaps %xmm1,%xmm9 movaps %xmm2,%xmm10 movaps %xmm5,%xmm11 ## type *=2 shll %r12d shll %r13d shll %r14d shll %r15d ## square it mulps %xmm1,%xmm1 mulps %xmm2,%xmm2 mulps %xmm5,%xmm5 addps %xmm2,%xmm1 addps %xmm5,%xmm1 ## rsq in xmm1 ## 2*type*ntia movl nb010_ntia(%rsp),%edi addl %edi,%r12d addl %edi,%r13d addl %edi,%r14d addl %edi,%r15d movq nb010_vdwparam(%rbp),%rsi ## xmm0=c6 ## xmm3=c12 rcpps %xmm1,%xmm5 ## 1/x lookup seed in xmm5 movaps nb010_two(%rsp),%xmm6 mulps %xmm5,%xmm1 ## load c6/c12 movlps (%rsi,%r12,4),%xmm7 movlps (%rsi,%r14,4),%xmm8 subps %xmm1,%xmm6 mulps %xmm5,%xmm6 ## xmm6=rinvsq movaps %xmm6,%xmm4 ## rinvsq movhps (%rsi,%r13,4),%xmm7 movhps (%rsi,%r15,4),%xmm8 movaps %xmm6,%xmm1 mulps %xmm6,%xmm1 ## rinv4 mulps %xmm6,%xmm1 ## rinv6 movaps %xmm1,%xmm2 mulps %xmm2,%xmm2 ## xmm2=rinv12 ## shuffle c6/c12 movaps %xmm7,%xmm5 shufps $136,%xmm8,%xmm5 ## 10001000 shufps $221,%xmm8,%xmm7 ## 11011101 movq nb010_faction(%rbp),%rsi mulps %xmm5,%xmm1 ## c6*rinv6 mulps %xmm7,%xmm2 ## c12*rinv12 movaps %xmm2,%xmm5 subps %xmm1,%xmm5 ## Vvdw=Vvdw12-Vvdw6 mulps nb010_six(%rsp),%xmm1 mulps nb010_twelve(%rsp),%xmm2 subps %xmm1,%xmm2 mulps %xmm2,%xmm4 ## xmm4=total fscal ## the fj's - start by accumulating x & y forces from memory movlps (%rsi,%r8,4),%xmm0 ## x1 y1 - - movlps (%rsi,%r10,4),%xmm1 ## x3 y3 - - movhps (%rsi,%r9,4),%xmm0 ## x1 y1 x2 y2 movhps (%rsi,%r11,4),%xmm1 ## x3 y3 x4 y4 ## add potential to Vvdwtot (sum in xmm12) addps %xmm5,%xmm12 ## 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 ## permute local forces movaps %xmm9,%xmm8 unpcklps %xmm10,%xmm9 ## x1 y1 x2 y2 unpckhps %xmm10,%xmm8 ## x3 y3 x4 y4 ## xmm11: fjz1 fjz2 fjz3 fjz4 pshufd $1,%xmm11,%xmm5 ## fjz2 - - - movhlps %xmm11,%xmm4 ## fjz3 - - - pshufd $3,%xmm11,%xmm3 ## fjz4 - - - ## update fjx and fjy addps %xmm9,%xmm0 addps %xmm8,%xmm1 movlps %xmm0,(%rsi,%r8,4) movlps %xmm1,(%rsi,%r10,4) movhps %xmm0,(%rsi,%r9,4) movhps %xmm1,(%rsi,%r11,4) addss 8(%rsi,%r8,4),%xmm11 addss 8(%rsi,%r9,4),%xmm5 addss 8(%rsi,%r10,4),%xmm4 addss 8(%rsi,%r11,4),%xmm3 movss %xmm11,8(%rsi,%r8,4) movss %xmm5,8(%rsi,%r9,4) movss %xmm4,8(%rsi,%r10,4) movss %xmm3,8(%rsi,%r11,4) ## should we do one more iteration? subl $4,nb010_innerk(%rsp) jl _nb_kernel010_x86_64_sse.nb010_finish_inner jmp _nb_kernel010_x86_64_sse.nb010_unroll_loop_nb_kernel010_x86_64_sse.nb010_finish_inner: ## check if at least two particles remain addl $4,nb010_innerk(%rsp) movl nb010_innerk(%rsp),%edx andl $2,%edx jnz _nb_kernel010_x86_64_sse.nb010_dopair jmp _nb_kernel010_x86_64_sse.nb010_checksingle_nb_kernel010_x86_64_sse.nb010_dopair: ## twice-unrolled innerloop here movq nb010_innerjjnr(%rsp),%rdx ## pointer to jjnr[k] movl (%rdx),%eax movl 4(%rdx),%ebx addq $8,nb010_innerjjnr(%rsp) ## advance pointer (unrolled 2) movq nb010_type(%rbp),%rsi movl (%rsi,%rax,4),%r12d movl (%rsi,%rbx,4),%r13d shll %r12d shll %r13d movl nb010_ntia(%rsp),%edi addl %edi,%r12d addl %edi,%r13d movq nb010_vdwparam(%rbp),%rsi movlps (%rsi,%r12,4),%xmm3 movhps (%rsi,%r13,4),%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?