1278 lines
37 KiB
Plaintext
1278 lines
37 KiB
Plaintext
.version 2.3
|
|
.target sm_20
|
|
.address_size 64
|
|
// compiled with /usr/local/cuda/open64/lib//be
|
|
// nvopencc 4.0 built on 2011-05-12
|
|
|
|
//-----------------------------------------------------------
|
|
// Compiling /tmp/tmpxft_0000c070_00000000-9_cmmc_long_gpu_kernel.cpp3.i (/home/sjplimp/ccBI#.PEOSwN)
|
|
//-----------------------------------------------------------
|
|
|
|
//-----------------------------------------------------------
|
|
// Options:
|
|
//-----------------------------------------------------------
|
|
// Target:ptx, ISA:sm_20, Endian:little, Pointer Size:64
|
|
// -O3 (Optimization level)
|
|
// -g0 (Debug level)
|
|
// -m2 (Report advisories)
|
|
//-----------------------------------------------------------
|
|
|
|
.file 1 "<command-line>"
|
|
.file 2 "/tmp/tmpxft_0000c070_00000000-8_cmmc_long_gpu_kernel.cudafe2.gpu"
|
|
.file 3 "/usr/lib/gcc/x86_64-redhat-linux/4.4.5/include/stddef.h"
|
|
.file 4 "/usr/local/cuda/include/crt/device_runtime.h"
|
|
.file 5 "/usr/local/cuda/include/host_defines.h"
|
|
.file 6 "/usr/local/cuda/include/builtin_types.h"
|
|
.file 7 "/usr/local/cuda/include/device_types.h"
|
|
.file 8 "/usr/local/cuda/include/driver_types.h"
|
|
.file 9 "/usr/local/cuda/include/surface_types.h"
|
|
.file 10 "/usr/local/cuda/include/texture_types.h"
|
|
.file 11 "/usr/local/cuda/include/vector_types.h"
|
|
.file 12 "/usr/local/cuda/include/device_launch_parameters.h"
|
|
.file 13 "/usr/local/cuda/include/crt/storage_class.h"
|
|
.file 14 "/usr/include/bits/types.h"
|
|
.file 15 "/usr/include/time.h"
|
|
.file 16 "cmmc_long_gpu_kernel.cu"
|
|
.file 17 "/usr/local/cuda/include/common_functions.h"
|
|
.file 18 "/usr/local/cuda/include/math_functions.h"
|
|
.file 19 "/usr/local/cuda/include/math_constants.h"
|
|
.file 20 "/usr/local/cuda/include/device_functions.h"
|
|
.file 21 "/usr/local/cuda/include/sm_11_atomic_functions.h"
|
|
.file 22 "/usr/local/cuda/include/sm_12_atomic_functions.h"
|
|
.file 23 "/usr/local/cuda/include/sm_13_double_functions.h"
|
|
.file 24 "/usr/local/cuda/include/sm_20_atomic_functions.h"
|
|
.file 25 "/usr/local/cuda/include/sm_20_intrinsics.h"
|
|
.file 26 "/usr/local/cuda/include/surface_functions.h"
|
|
.file 27 "/usr/local/cuda/include/texture_fetch_functions.h"
|
|
.file 28 "/usr/local/cuda/include/math_functions_dbl_ptx3.h"
|
|
|
|
.global .texref pos_tex;
|
|
.global .texref q_tex;
|
|
|
|
.entry kernel_pair (
|
|
.param .u64 __cudaparm_kernel_pair_x_,
|
|
.param .u64 __cudaparm_kernel_pair_lj1,
|
|
.param .u64 __cudaparm_kernel_pair_lj3,
|
|
.param .s32 __cudaparm_kernel_pair_lj_types,
|
|
.param .u64 __cudaparm_kernel_pair_sp_lj_in,
|
|
.param .u64 __cudaparm_kernel_pair_dev_nbor,
|
|
.param .u64 __cudaparm_kernel_pair_dev_packed,
|
|
.param .u64 __cudaparm_kernel_pair_ans,
|
|
.param .u64 __cudaparm_kernel_pair_engv,
|
|
.param .s32 __cudaparm_kernel_pair_eflag,
|
|
.param .s32 __cudaparm_kernel_pair_vflag,
|
|
.param .s32 __cudaparm_kernel_pair_inum,
|
|
.param .s32 __cudaparm_kernel_pair_nbor_pitch,
|
|
.param .u64 __cudaparm_kernel_pair_q_,
|
|
.param .f32 __cudaparm_kernel_pair_cut_coulsq,
|
|
.param .f32 __cudaparm_kernel_pair_qqrd2e,
|
|
.param .f32 __cudaparm_kernel_pair_g_ewald,
|
|
.param .s32 __cudaparm_kernel_pair_t_per_atom)
|
|
{
|
|
.reg .u32 %r<86>;
|
|
.reg .u64 %rd<65>;
|
|
.reg .f32 %f<175>;
|
|
.reg .pred %p<23>;
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32498_33_non_const_sp_lj112[32];
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32619_35_non_const_red_acc144[3072];
|
|
// __cuda_local_var_32510_10_non_const_f = 64
|
|
// __cuda_local_var_32514_9_non_const_virial = 16
|
|
// __cuda_local_var_32561_43_non_const_inv1 = 40
|
|
// __cuda_local_var_32561_49_non_const_inv2 = 44
|
|
// __cuda_local_var_32561_55_non_const_prefactor = 52
|
|
// __cuda_local_var_32561_66_non_const__erfc = 48
|
|
.loc 16 108 0
|
|
$LDWbegin_kernel_pair:
|
|
.loc 16 115 0
|
|
ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];
|
|
ldu.global.f32 %f1, [%rd1+0];
|
|
.loc 16 116 0
|
|
ld.global.f32 %f2, [%rd1+4];
|
|
.loc 16 117 0
|
|
ld.global.f32 %f3, [%rd1+8];
|
|
.loc 16 118 0
|
|
ld.global.f32 %f4, [%rd1+12];
|
|
st.shared.v4.f32 [__cuda___cuda_local_var_32498_33_non_const_sp_lj112+0], {%f1,%f2,%f3,%f4};
|
|
.loc 16 119 0
|
|
ld.global.f32 %f5, [%rd1+16];
|
|
.loc 16 120 0
|
|
ld.global.f32 %f6, [%rd1+20];
|
|
.loc 16 121 0
|
|
ld.global.f32 %f7, [%rd1+24];
|
|
.loc 16 122 0
|
|
ld.global.f32 %f8, [%rd1+28];
|
|
st.shared.v4.f32 [__cuda___cuda_local_var_32498_33_non_const_sp_lj112+16], {%f5,%f6,%f7,%f8};
|
|
.loc 16 132 0
|
|
mov.f32 %f9, 0f00000000; // 0
|
|
mov.f32 %f10, %f9;
|
|
mov.f32 %f11, 0f00000000; // 0
|
|
mov.f32 %f12, %f11;
|
|
mov.f32 %f13, 0f00000000; // 0
|
|
mov.f32 %f14, %f13;
|
|
mov.f32 %f15, 0f00000000; // 0
|
|
mov.f32 %f16, %f15;
|
|
mov.f32 %f17, 0f00000000; // 0
|
|
mov.f32 %f18, %f17;
|
|
mov.f32 %f19, 0f00000000; // 0
|
|
mov.f32 %f20, %f19;
|
|
ld.param.s32 %r1, [__cudaparm_kernel_pair_t_per_atom];
|
|
cvt.s32.u32 %r2, %tid.x;
|
|
div.s32 %r3, %r2, %r1;
|
|
cvt.s32.u32 %r4, %ntid.x;
|
|
div.s32 %r5, %r4, %r1;
|
|
rem.s32 %r6, %r2, %r1;
|
|
cvt.s32.u32 %r7, %ctaid.x;
|
|
mul.lo.s32 %r8, %r7, %r5;
|
|
add.s32 %r9, %r3, %r8;
|
|
ld.param.s32 %r10, [__cudaparm_kernel_pair_inum];
|
|
setp.lt.s32 %p1, %r9, %r10;
|
|
@!%p1 bra $Lt_0_23810;
|
|
.loc 16 136 0
|
|
cvt.s64.s32 %rd2, %r9;
|
|
mul.wide.s32 %rd3, %r9, 4;
|
|
ld.param.u64 %rd4, [__cudaparm_kernel_pair_dev_nbor];
|
|
add.u64 %rd5, %rd3, %rd4;
|
|
ld.global.s32 %r11, [%rd5+0];
|
|
.loc 16 138 0
|
|
ld.param.s32 %r12, [__cudaparm_kernel_pair_nbor_pitch];
|
|
cvt.s64.s32 %rd6, %r12;
|
|
mul.wide.s32 %rd7, %r12, 4;
|
|
add.u64 %rd8, %rd7, %rd5;
|
|
ld.global.s32 %r13, [%rd8+0];
|
|
add.u64 %rd9, %rd7, %rd8;
|
|
ld.param.u64 %rd10, [__cudaparm_kernel_pair_dev_packed];
|
|
setp.ne.u64 %p2, %rd10, %rd4;
|
|
@%p2 bra $Lt_0_24322;
|
|
.loc 16 144 0
|
|
cvt.s32.s64 %r14, %rd6;
|
|
mul.lo.s32 %r15, %r14, %r13;
|
|
cvt.s64.s32 %rd11, %r15;
|
|
mul.wide.s32 %rd12, %r15, 4;
|
|
add.u64 %rd13, %rd9, %rd12;
|
|
.loc 16 145 0
|
|
mul.lo.s32 %r16, %r6, %r14;
|
|
cvt.s64.s32 %rd14, %r16;
|
|
mul.wide.s32 %rd15, %r16, 4;
|
|
add.u64 %rd16, %rd9, %rd15;
|
|
.loc 16 146 0
|
|
mul.lo.s32 %r17, %r14, %r1;
|
|
bra.uni $Lt_0_24066;
|
|
$Lt_0_24322:
|
|
.loc 16 148 0
|
|
ld.global.s32 %r18, [%rd9+0];
|
|
cvt.s64.s32 %rd17, %r18;
|
|
mul.wide.s32 %rd18, %r18, 4;
|
|
add.u64 %rd19, %rd10, %rd18;
|
|
.loc 16 149 0
|
|
cvt.s64.s32 %rd20, %r13;
|
|
mul.wide.s32 %rd21, %r13, 4;
|
|
add.u64 %rd13, %rd19, %rd21;
|
|
.loc 16 150 0
|
|
mov.s32 %r17, %r1;
|
|
.loc 16 151 0
|
|
cvt.s64.s32 %rd22, %r6;
|
|
mul.wide.s32 %rd23, %r6, 4;
|
|
add.u64 %rd16, %rd19, %rd23;
|
|
$Lt_0_24066:
|
|
.loc 16 154 0
|
|
mov.u32 %r19, %r11;
|
|
mov.s32 %r20, 0;
|
|
mov.u32 %r21, %r20;
|
|
mov.s32 %r22, 0;
|
|
mov.u32 %r23, %r22;
|
|
mov.s32 %r24, 0;
|
|
mov.u32 %r25, %r24;
|
|
tex.1d.v4.f32.s32 {%f21,%f22,%f23,%f24},[pos_tex,{%r19,%r21,%r23,%r25}];
|
|
mov.f32 %f25, %f21;
|
|
mov.f32 %f26, %f22;
|
|
mov.f32 %f27, %f23;
|
|
mov.f32 %f28, %f24;
|
|
.loc 16 155 0
|
|
mov.u32 %r26, %r11;
|
|
mov.s32 %r27, 0;
|
|
mov.u32 %r28, %r27;
|
|
mov.s32 %r29, 0;
|
|
mov.u32 %r30, %r29;
|
|
mov.s32 %r31, 0;
|
|
mov.u32 %r32, %r31;
|
|
tex.1d.v4.f32.s32 {%f29,%f30,%f31,%f32},[q_tex,{%r26,%r28,%r30,%r32}];
|
|
mov.f32 %f33, %f29;
|
|
setp.ge.u64 %p3, %rd16, %rd13;
|
|
@%p3 bra $Lt_0_35330;
|
|
cvt.rzi.ftz.s32.f32 %r33, %f28;
|
|
cvt.s64.s32 %rd24, %r17;
|
|
ld.param.s32 %r34, [__cudaparm_kernel_pair_lj_types];
|
|
mul.lo.s32 %r35, %r34, %r33;
|
|
ld.param.u64 %rd25, [__cudaparm_kernel_pair_lj1];
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
mov.f32 %f35, 0f00000000; // 0
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
mov.u64 %rd26, __cuda___cuda_local_var_32498_33_non_const_sp_lj112;
|
|
$Lt_0_25090:
|
|
//<loop> Loop body line 155, nesting depth: 1, estimated iterations: unknown
|
|
.loc 16 159 0
|
|
ld.global.s32 %r36, [%rd16+0];
|
|
.loc 16 162 0
|
|
shr.s32 %r37, %r36, 30;
|
|
and.b32 %r38, %r37, 3;
|
|
cvt.s64.s32 %rd27, %r38;
|
|
mul.wide.s32 %rd28, %r38, 4;
|
|
add.u64 %rd29, %rd26, %rd28;
|
|
ld.shared.f32 %f39, [%rd29+0];
|
|
.loc 16 163 0
|
|
mov.f32 %f40, 0f3f800000; // 1
|
|
ld.shared.f32 %f41, [%rd29+16];
|
|
sub.ftz.f32 %f42, %f40, %f41;
|
|
.loc 16 166 0
|
|
and.b32 %r39, %r36, 1073741823;
|
|
mov.u32 %r40, %r39;
|
|
mov.s32 %r41, 0;
|
|
mov.u32 %r42, %r41;
|
|
mov.s32 %r43, 0;
|
|
mov.u32 %r44, %r43;
|
|
mov.s32 %r45, 0;
|
|
mov.u32 %r46, %r45;
|
|
tex.1d.v4.f32.s32 {%f43,%f44,%f45,%f46},[pos_tex,{%r40,%r42,%r44,%r46}];
|
|
mov.f32 %f47, %f43;
|
|
mov.f32 %f48, %f44;
|
|
mov.f32 %f49, %f45;
|
|
mov.f32 %f50, %f46;
|
|
cvt.rzi.ftz.s32.f32 %r47, %f50;
|
|
sub.ftz.f32 %f51, %f26, %f48;
|
|
sub.ftz.f32 %f52, %f25, %f47;
|
|
sub.ftz.f32 %f53, %f27, %f49;
|
|
mul.ftz.f32 %f54, %f51, %f51;
|
|
fma.rn.ftz.f32 %f55, %f52, %f52, %f54;
|
|
fma.rn.ftz.f32 %f56, %f53, %f53, %f55;
|
|
add.s32 %r48, %r47, %r35;
|
|
cvt.s64.s32 %rd30, %r48;
|
|
mul.wide.s32 %rd31, %r48, 16;
|
|
add.u64 %rd32, %rd31, %rd25;
|
|
ld.global.f32 %f57, [%rd32+0];
|
|
setp.gt.ftz.f32 %p4, %f57, %f56;
|
|
@!%p4 bra $Lt_0_28930;
|
|
rcp.approx.ftz.f32 %f58, %f56;
|
|
ld.global.f32 %f59, [%rd32+4];
|
|
setp.lt.ftz.f32 %p5, %f56, %f59;
|
|
@!%p5 bra $Lt_0_26114;
|
|
ld.param.u64 %rd33, [__cudaparm_kernel_pair_lj3];
|
|
add.u64 %rd34, %rd33, %rd31;
|
|
ld.global.f32 %f60, [%rd34+0];
|
|
mov.f32 %f61, 0f40000000; // 2
|
|
setp.eq.ftz.f32 %p6, %f60, %f61;
|
|
@!%p6 bra $Lt_0_26626;
|
|
.loc 16 182 0
|
|
mul.ftz.f32 %f62, %f58, %f58;
|
|
mov.f32 %f63, %f62;
|
|
mov.f32 %f64, %f63;
|
|
.loc 16 183 0
|
|
mul.ftz.f32 %f65, %f62, %f62;
|
|
mov.f32 %f66, %f65;
|
|
bra.uni $Lt_0_26882;
|
|
$Lt_0_26626:
|
|
mov.f32 %f67, 0f3f800000; // 1
|
|
setp.eq.ftz.f32 %p7, %f60, %f67;
|
|
@!%p7 bra $Lt_0_27138;
|
|
.loc 16 185 0
|
|
sqrt.approx.ftz.f32 %f68, %f58;
|
|
mul.ftz.f32 %f69, %f58, %f68;
|
|
mov.f32 %f65, %f69;
|
|
mov.f32 %f66, %f65;
|
|
.loc 16 186 0
|
|
mul.ftz.f32 %f63, %f69, %f69;
|
|
mov.f32 %f64, %f63;
|
|
bra.uni $Lt_0_26882;
|
|
$Lt_0_27138:
|
|
.loc 16 188 0
|
|
mul.ftz.f32 %f70, %f58, %f58;
|
|
mul.ftz.f32 %f71, %f58, %f70;
|
|
mov.f32 %f63, %f71;
|
|
mov.f32 %f64, %f63;
|
|
.loc 16 189 0
|
|
mov.f32 %f65, %f71;
|
|
mov.f32 %f66, %f65;
|
|
$Lt_0_26882:
|
|
$Lt_0_26370:
|
|
.loc 16 191 0
|
|
mul.ftz.f32 %f72, %f39, %f63;
|
|
ld.global.v2.f32 {%f73,%f74}, [%rd32+8];
|
|
mul.ftz.f32 %f75, %f73, %f65;
|
|
sub.ftz.f32 %f76, %f75, %f74;
|
|
mul.ftz.f32 %f77, %f72, %f76;
|
|
bra.uni $Lt_0_25858;
|
|
$Lt_0_26114:
|
|
.loc 16 193 0
|
|
mov.f32 %f77, 0f00000000; // 0
|
|
$Lt_0_25858:
|
|
ld.param.f32 %f78, [__cudaparm_kernel_pair_cut_coulsq];
|
|
setp.gt.ftz.f32 %p8, %f78, %f56;
|
|
@!%p8 bra $Lt_0_27650;
|
|
.loc 16 200 0
|
|
sqrt.approx.ftz.f32 %f79, %f56;
|
|
ld.param.f32 %f80, [__cudaparm_kernel_pair_g_ewald];
|
|
mul.ftz.f32 %f81, %f80, %f79;
|
|
mul.ftz.f32 %f82, %f81, %f81;
|
|
mov.f32 %f83, 0f3f800000; // 1
|
|
mov.f32 %f84, 0f3ea7ba05; // 0.327591
|
|
fma.rn.ftz.f32 %f85, %f84, %f81, %f83;
|
|
neg.ftz.f32 %f86, %f82;
|
|
rcp.approx.ftz.f32 %f87, %f85;
|
|
mov.f32 %f88, 0f3fb8aa3b; // 1.4427
|
|
mul.ftz.f32 %f89, %f86, %f88;
|
|
ex2.approx.ftz.f32 %f90, %f89;
|
|
mov.f32 %f91, 0f3e827906; // 0.25483
|
|
mov.f32 %f92, 0fbe91a98e; // -0.284497
|
|
mov.f32 %f93, 0f3fb5f0e3; // 1.42141
|
|
mov.f32 %f94, 0fbfba00e3; // -1.45315
|
|
mov.f32 %f95, 0f3f87dc22; // 1.06141
|
|
fma.rn.ftz.f32 %f96, %f95, %f87, %f94;
|
|
fma.rn.ftz.f32 %f97, %f87, %f96, %f93;
|
|
fma.rn.ftz.f32 %f98, %f87, %f97, %f92;
|
|
fma.rn.ftz.f32 %f99, %f87, %f98, %f91;
|
|
mul.ftz.f32 %f100, %f87, %f99;
|
|
mul.ftz.f32 %f101, %f90, %f100;
|
|
mov.f32 %f102, %f101;
|
|
.loc 16 201 0
|
|
mov.u32 %r49, %r39;
|
|
mov.s32 %r50, 0;
|
|
mov.u32 %r51, %r50;
|
|
mov.s32 %r52, 0;
|
|
mov.u32 %r53, %r52;
|
|
mov.s32 %r54, 0;
|
|
mov.u32 %r55, %r54;
|
|
tex.1d.v4.f32.s32 {%f103,%f104,%f105,%f106},[q_tex,{%r49,%r51,%r53,%r55}];
|
|
mov.f32 %f107, %f103;
|
|
ld.param.f32 %f108, [__cudaparm_kernel_pair_qqrd2e];
|
|
mul.ftz.f32 %f109, %f108, %f33;
|
|
mul.ftz.f32 %f110, %f109, %f107;
|
|
div.approx.ftz.f32 %f111, %f110, %f79;
|
|
mov.f32 %f112, %f111;
|
|
.loc 16 202 0
|
|
mov.f32 %f113, 0f3f906ebb; // 1.12838
|
|
mul.ftz.f32 %f114, %f81, %f113;
|
|
fma.rn.ftz.f32 %f115, %f90, %f114, %f101;
|
|
sub.ftz.f32 %f116, %f115, %f42;
|
|
mul.ftz.f32 %f117, %f111, %f116;
|
|
bra.uni $Lt_0_27394;
|
|
$Lt_0_27650:
|
|
.loc 16 204 0
|
|
mov.f32 %f117, 0f00000000; // 0
|
|
$Lt_0_27394:
|
|
.loc 16 208 0
|
|
add.ftz.f32 %f118, %f117, %f77;
|
|
mul.ftz.f32 %f119, %f118, %f58;
|
|
fma.rn.ftz.f32 %f36, %f52, %f119, %f36;
|
|
.loc 16 209 0
|
|
fma.rn.ftz.f32 %f35, %f51, %f119, %f35;
|
|
.loc 16 210 0
|
|
fma.rn.ftz.f32 %f34, %f53, %f119, %f34;
|
|
ld.param.s32 %r56, [__cudaparm_kernel_pair_eflag];
|
|
mov.u32 %r57, 0;
|
|
setp.le.s32 %p9, %r56, %r57;
|
|
@%p9 bra $Lt_0_28418;
|
|
.loc 16 213 0
|
|
mov.f32 %f120, %f112;
|
|
mov.f32 %f121, %f102;
|
|
sub.ftz.f32 %f122, %f121, %f42;
|
|
fma.rn.ftz.f32 %f123, %f120, %f122, %f37;
|
|
selp.f32 %f37, %f123, %f37, %p8;
|
|
@!%p5 bra $Lt_0_28418;
|
|
.loc 16 216 0
|
|
ld.param.u64 %rd35, [__cudaparm_kernel_pair_lj3];
|
|
add.u64 %rd36, %rd35, %rd31;
|
|
ld.global.v4.f32 {_,%f124,%f125,%f126}, [%rd36+0];
|
|
mov.f32 %f127, %f64;
|
|
mul.ftz.f32 %f128, %f127, %f39;
|
|
mov.f32 %f129, %f66;
|
|
mul.ftz.f32 %f130, %f124, %f129;
|
|
sub.ftz.f32 %f131, %f130, %f125;
|
|
mul.ftz.f32 %f132, %f128, %f131;
|
|
sub.ftz.f32 %f133, %f132, %f126;
|
|
add.ftz.f32 %f38, %f38, %f133;
|
|
$Lt_0_28418:
|
|
$Lt_0_27906:
|
|
ld.param.s32 %r58, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r59, 0;
|
|
setp.le.s32 %p10, %r58, %r59;
|
|
@%p10 bra $Lt_0_28930;
|
|
.loc 16 221 0
|
|
mov.f32 %f134, %f10;
|
|
mul.ftz.f32 %f135, %f52, %f52;
|
|
fma.rn.ftz.f32 %f136, %f119, %f135, %f134;
|
|
mov.f32 %f10, %f136;
|
|
.loc 16 222 0
|
|
mov.f32 %f137, %f12;
|
|
fma.rn.ftz.f32 %f138, %f119, %f54, %f137;
|
|
mov.f32 %f12, %f138;
|
|
.loc 16 223 0
|
|
mov.f32 %f139, %f14;
|
|
mul.ftz.f32 %f140, %f53, %f53;
|
|
fma.rn.ftz.f32 %f141, %f119, %f140, %f139;
|
|
mov.f32 %f14, %f141;
|
|
.loc 16 224 0
|
|
mov.f32 %f142, %f16;
|
|
mul.ftz.f32 %f143, %f51, %f52;
|
|
fma.rn.ftz.f32 %f144, %f119, %f143, %f142;
|
|
mov.f32 %f16, %f144;
|
|
.loc 16 225 0
|
|
mov.f32 %f145, %f18;
|
|
mul.ftz.f32 %f146, %f52, %f53;
|
|
fma.rn.ftz.f32 %f147, %f119, %f146, %f145;
|
|
mov.f32 %f18, %f147;
|
|
.loc 16 226 0
|
|
mul.ftz.f32 %f148, %f51, %f53;
|
|
fma.rn.ftz.f32 %f19, %f119, %f148, %f19;
|
|
mov.f32 %f20, %f19;
|
|
$Lt_0_28930:
|
|
$Lt_0_25346:
|
|
.loc 16 158 0
|
|
mul.lo.u64 %rd37, %rd24, 4;
|
|
add.u64 %rd16, %rd16, %rd37;
|
|
setp.lt.u64 %p11, %rd16, %rd13;
|
|
@%p11 bra $Lt_0_25090;
|
|
bra.uni $Lt_0_23554;
|
|
$Lt_0_35330:
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
mov.f32 %f35, 0f00000000; // 0
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
bra.uni $Lt_0_23554;
|
|
$Lt_0_23810:
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
mov.f32 %f35, 0f00000000; // 0
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
$Lt_0_23554:
|
|
mov.u32 %r60, 1;
|
|
setp.le.s32 %p12, %r1, %r60;
|
|
@%p12 bra $Lt_0_31746;
|
|
.loc 16 237 0
|
|
mov.u64 %rd38, __cuda___cuda_local_var_32619_35_non_const_red_acc144;
|
|
cvt.s64.s32 %rd39, %r2;
|
|
mul.wide.s32 %rd40, %r2, 4;
|
|
add.u64 %rd41, %rd38, %rd40;
|
|
mov.f32 %f149, %f36;
|
|
st.shared.f32 [%rd41+0], %f149;
|
|
.loc 16 238 0
|
|
mov.f32 %f150, %f35;
|
|
st.shared.f32 [%rd41+512], %f150;
|
|
.loc 16 239 0
|
|
mov.f32 %f151, %f34;
|
|
st.shared.f32 [%rd41+1024], %f151;
|
|
.loc 16 240 0
|
|
mov.f32 %f152, %f38;
|
|
st.shared.f32 [%rd41+1536], %f152;
|
|
.loc 16 241 0
|
|
mov.f32 %f153, %f37;
|
|
st.shared.f32 [%rd41+2048], %f153;
|
|
.loc 16 243 0
|
|
shr.s32 %r61, %r1, 31;
|
|
mov.s32 %r62, 1;
|
|
and.b32 %r63, %r61, %r62;
|
|
add.s32 %r64, %r63, %r1;
|
|
shr.s32 %r65, %r64, 1;
|
|
mov.s32 %r66, %r65;
|
|
mov.u32 %r67, 0;
|
|
setp.ne.u32 %p13, %r65, %r67;
|
|
@!%p13 bra $Lt_0_30210;
|
|
$Lt_0_30722:
|
|
setp.ge.u32 %p14, %r6, %r66;
|
|
@%p14 bra $Lt_0_30978;
|
|
.loc 16 246 0
|
|
add.u32 %r68, %r2, %r66;
|
|
cvt.u64.u32 %rd42, %r68;
|
|
mul.wide.u32 %rd43, %r68, 4;
|
|
add.u64 %rd44, %rd38, %rd43;
|
|
ld.shared.f32 %f154, [%rd44+0];
|
|
add.ftz.f32 %f149, %f154, %f149;
|
|
st.shared.f32 [%rd41+0], %f149;
|
|
ld.shared.f32 %f155, [%rd44+512];
|
|
add.ftz.f32 %f150, %f155, %f150;
|
|
st.shared.f32 [%rd41+512], %f150;
|
|
ld.shared.f32 %f156, [%rd44+1024];
|
|
add.ftz.f32 %f151, %f156, %f151;
|
|
st.shared.f32 [%rd41+1024], %f151;
|
|
ld.shared.f32 %f157, [%rd44+1536];
|
|
add.ftz.f32 %f152, %f157, %f152;
|
|
st.shared.f32 [%rd41+1536], %f152;
|
|
ld.shared.f32 %f158, [%rd44+2048];
|
|
add.ftz.f32 %f153, %f158, %f153;
|
|
st.shared.f32 [%rd41+2048], %f153;
|
|
$Lt_0_30978:
|
|
.loc 16 243 0
|
|
shr.u32 %r66, %r66, 1;
|
|
mov.u32 %r69, 0;
|
|
setp.ne.u32 %p15, %r66, %r69;
|
|
@%p15 bra $Lt_0_30722;
|
|
$Lt_0_30210:
|
|
.loc 16 250 0
|
|
mov.f32 %f36, %f149;
|
|
.loc 16 251 0
|
|
mov.f32 %f35, %f150;
|
|
.loc 16 252 0
|
|
mov.f32 %f34, %f151;
|
|
.loc 16 253 0
|
|
mov.f32 %f38, %f152;
|
|
.loc 16 254 0
|
|
mov.f32 %f37, %f153;
|
|
ld.param.s32 %r70, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r71, 0;
|
|
setp.le.s32 %p16, %r70, %r71;
|
|
@%p16 bra $Lt_0_31746;
|
|
.loc 16 258 0
|
|
mov.f32 %f149, %f10;
|
|
st.shared.f32 [%rd41+0], %f149;
|
|
mov.f32 %f150, %f12;
|
|
st.shared.f32 [%rd41+512], %f150;
|
|
mov.f32 %f151, %f14;
|
|
st.shared.f32 [%rd41+1024], %f151;
|
|
mov.f32 %f152, %f16;
|
|
st.shared.f32 [%rd41+1536], %f152;
|
|
mov.f32 %f153, %f18;
|
|
st.shared.f32 [%rd41+2048], %f153;
|
|
mov.f32 %f159, %f20;
|
|
st.shared.f32 [%rd41+2560], %f159;
|
|
.loc 16 260 0
|
|
mov.s32 %r72, %r65;
|
|
@!%p13 bra $Lt_0_32258;
|
|
$Lt_0_32770:
|
|
setp.ge.u32 %p17, %r6, %r72;
|
|
@%p17 bra $Lt_0_33026;
|
|
.loc 16 263 0
|
|
add.u32 %r73, %r2, %r72;
|
|
cvt.u64.u32 %rd45, %r73;
|
|
mul.wide.u32 %rd46, %r73, 4;
|
|
add.u64 %rd47, %rd38, %rd46;
|
|
ld.shared.f32 %f160, [%rd47+0];
|
|
add.ftz.f32 %f149, %f160, %f149;
|
|
st.shared.f32 [%rd41+0], %f149;
|
|
ld.shared.f32 %f161, [%rd47+512];
|
|
add.ftz.f32 %f150, %f161, %f150;
|
|
st.shared.f32 [%rd41+512], %f150;
|
|
ld.shared.f32 %f162, [%rd47+1024];
|
|
add.ftz.f32 %f151, %f162, %f151;
|
|
st.shared.f32 [%rd41+1024], %f151;
|
|
ld.shared.f32 %f163, [%rd47+1536];
|
|
add.ftz.f32 %f152, %f163, %f152;
|
|
st.shared.f32 [%rd41+1536], %f152;
|
|
ld.shared.f32 %f164, [%rd47+2048];
|
|
add.ftz.f32 %f153, %f164, %f153;
|
|
st.shared.f32 [%rd41+2048], %f153;
|
|
ld.shared.f32 %f165, [%rd47+2560];
|
|
add.ftz.f32 %f159, %f165, %f159;
|
|
st.shared.f32 [%rd41+2560], %f159;
|
|
$Lt_0_33026:
|
|
.loc 16 260 0
|
|
shr.u32 %r72, %r72, 1;
|
|
mov.u32 %r74, 0;
|
|
setp.ne.u32 %p18, %r72, %r74;
|
|
@%p18 bra $Lt_0_32770;
|
|
$Lt_0_32258:
|
|
.loc 16 268 0
|
|
mov.f32 %f10, %f149;
|
|
mov.f32 %f12, %f150;
|
|
mov.f32 %f14, %f151;
|
|
mov.f32 %f16, %f152;
|
|
mov.f32 %f18, %f153;
|
|
mov.f32 %f20, %f159;
|
|
$Lt_0_31746:
|
|
$Lt_0_29698:
|
|
selp.s32 %r75, 1, 0, %p1;
|
|
mov.s32 %r76, 0;
|
|
set.eq.u32.s32 %r77, %r6, %r76;
|
|
neg.s32 %r78, %r77;
|
|
and.b32 %r79, %r75, %r78;
|
|
mov.u32 %r80, 0;
|
|
setp.eq.s32 %p19, %r79, %r80;
|
|
@%p19 bra $Lt_0_33794;
|
|
.loc 16 274 0
|
|
cvt.s64.s32 %rd48, %r9;
|
|
ld.param.u64 %rd49, [__cudaparm_kernel_pair_engv];
|
|
mul.wide.s32 %rd50, %r9, 4;
|
|
add.u64 %rd51, %rd49, %rd50;
|
|
ld.param.s32 %r81, [__cudaparm_kernel_pair_eflag];
|
|
mov.u32 %r82, 0;
|
|
setp.le.s32 %p20, %r81, %r82;
|
|
@%p20 bra $Lt_0_34306;
|
|
.loc 16 276 0
|
|
st.global.f32 [%rd51+0], %f38;
|
|
.loc 16 277 0
|
|
cvt.s64.s32 %rd52, %r10;
|
|
mul.wide.s32 %rd53, %r10, 4;
|
|
add.u64 %rd54, %rd53, %rd51;
|
|
.loc 16 278 0
|
|
st.global.f32 [%rd54+0], %f37;
|
|
.loc 16 279 0
|
|
add.u64 %rd51, %rd53, %rd54;
|
|
$Lt_0_34306:
|
|
ld.param.s32 %r83, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r84, 0;
|
|
setp.le.s32 %p21, %r83, %r84;
|
|
@%p21 bra $Lt_0_34818;
|
|
.loc 16 283 0
|
|
mov.f32 %f166, %f10;
|
|
st.global.f32 [%rd51+0], %f166;
|
|
.loc 16 284 0
|
|
cvt.s64.s32 %rd55, %r10;
|
|
mul.wide.s32 %rd56, %r10, 4;
|
|
add.u64 %rd57, %rd56, %rd51;
|
|
.loc 16 283 0
|
|
mov.f32 %f167, %f12;
|
|
st.global.f32 [%rd57+0], %f167;
|
|
.loc 16 284 0
|
|
add.u64 %rd58, %rd56, %rd57;
|
|
.loc 16 283 0
|
|
mov.f32 %f168, %f14;
|
|
st.global.f32 [%rd58+0], %f168;
|
|
.loc 16 284 0
|
|
add.u64 %rd59, %rd56, %rd58;
|
|
.loc 16 283 0
|
|
mov.f32 %f169, %f16;
|
|
st.global.f32 [%rd59+0], %f169;
|
|
.loc 16 284 0
|
|
add.u64 %rd51, %rd56, %rd59;
|
|
.loc 16 283 0
|
|
mov.f32 %f170, %f18;
|
|
st.global.f32 [%rd51+0], %f170;
|
|
mov.f32 %f171, %f20;
|
|
add.u64 %rd60, %rd56, %rd51;
|
|
st.global.f32 [%rd60+0], %f171;
|
|
$Lt_0_34818:
|
|
.loc 16 287 0
|
|
ld.param.u64 %rd61, [__cudaparm_kernel_pair_ans];
|
|
mul.lo.u64 %rd62, %rd48, 16;
|
|
add.u64 %rd63, %rd61, %rd62;
|
|
mov.f32 %f172, %f173;
|
|
st.global.v4.f32 [%rd63+0], {%f36,%f35,%f34,%f172};
|
|
$Lt_0_33794:
|
|
.loc 16 289 0
|
|
exit;
|
|
$LDWend_kernel_pair:
|
|
} // kernel_pair
|
|
|
|
.entry kernel_pair_fast (
|
|
.param .u64 __cudaparm_kernel_pair_fast_x_,
|
|
.param .u64 __cudaparm_kernel_pair_fast_lj1_in,
|
|
.param .u64 __cudaparm_kernel_pair_fast_lj3_in,
|
|
.param .u64 __cudaparm_kernel_pair_fast_sp_lj_in,
|
|
.param .u64 __cudaparm_kernel_pair_fast_dev_nbor,
|
|
.param .u64 __cudaparm_kernel_pair_fast_dev_packed,
|
|
.param .u64 __cudaparm_kernel_pair_fast_ans,
|
|
.param .u64 __cudaparm_kernel_pair_fast_engv,
|
|
.param .s32 __cudaparm_kernel_pair_fast_eflag,
|
|
.param .s32 __cudaparm_kernel_pair_fast_vflag,
|
|
.param .s32 __cudaparm_kernel_pair_fast_inum,
|
|
.param .s32 __cudaparm_kernel_pair_fast_nbor_pitch,
|
|
.param .u64 __cudaparm_kernel_pair_fast_q_,
|
|
.param .f32 __cudaparm_kernel_pair_fast_cut_coulsq,
|
|
.param .f32 __cudaparm_kernel_pair_fast_qqrd2e,
|
|
.param .f32 __cudaparm_kernel_pair_fast_g_ewald,
|
|
.param .s32 __cudaparm_kernel_pair_fast_t_per_atom)
|
|
{
|
|
.reg .u32 %r<86>;
|
|
.reg .u64 %rd<76>;
|
|
.reg .f32 %f<178>;
|
|
.reg .pred %p<25>;
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32691_33_non_const_sp_lj3320[32];
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32689_34_non_const_lj13360[1936];
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32690_34_non_const_lj35296[1936];
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32812_35_non_const_red_acc7232[3072];
|
|
// __cuda_local_var_32701_10_non_const_f = 64
|
|
// __cuda_local_var_32705_9_non_const_virial = 16
|
|
// __cuda_local_var_32754_43_non_const_inv1 = 40
|
|
// __cuda_local_var_32754_49_non_const_inv2 = 44
|
|
// __cuda_local_var_32754_55_non_const_prefactor = 52
|
|
// __cuda_local_var_32754_66_non_const__erfc = 48
|
|
.loc 16 299 0
|
|
$LDWbegin_kernel_pair_fast:
|
|
cvt.s32.u32 %r1, %tid.x;
|
|
mov.u32 %r2, 7;
|
|
setp.gt.s32 %p1, %r1, %r2;
|
|
@%p1 bra $Lt_1_25090;
|
|
.loc 16 309 0
|
|
mov.u64 %rd1, __cuda___cuda_local_var_32691_33_non_const_sp_lj3320;
|
|
cvt.s64.s32 %rd2, %r1;
|
|
mul.wide.s32 %rd3, %r1, 4;
|
|
ld.param.u64 %rd4, [__cudaparm_kernel_pair_fast_sp_lj_in];
|
|
add.u64 %rd5, %rd4, %rd3;
|
|
ld.global.f32 %f1, [%rd5+0];
|
|
add.u64 %rd6, %rd3, %rd1;
|
|
st.shared.f32 [%rd6+0], %f1;
|
|
$Lt_1_25090:
|
|
mov.u64 %rd1, __cuda___cuda_local_var_32691_33_non_const_sp_lj3320;
|
|
mov.u32 %r3, 120;
|
|
setp.gt.s32 %p2, %r1, %r3;
|
|
@%p2 bra $Lt_1_25602;
|
|
.loc 16 311 0
|
|
mov.u64 %rd7, __cuda___cuda_local_var_32689_34_non_const_lj13360;
|
|
mov.u64 %rd8, __cuda___cuda_local_var_32690_34_non_const_lj35296;
|
|
cvt.s64.s32 %rd9, %r1;
|
|
mul.wide.s32 %rd10, %r1, 16;
|
|
ld.param.u64 %rd11, [__cudaparm_kernel_pair_fast_lj1_in];
|
|
add.u64 %rd12, %rd11, %rd10;
|
|
add.u64 %rd13, %rd10, %rd7;
|
|
ld.global.v4.f32 {%f2,%f3,%f4,%f5}, [%rd12+0];
|
|
st.shared.v4.f32 [%rd13+0], {%f2,%f3,%f4,%f5};
|
|
.loc 16 312 0
|
|
ld.param.u64 %rd14, [__cudaparm_kernel_pair_fast_lj3_in];
|
|
add.u64 %rd15, %rd14, %rd10;
|
|
add.u64 %rd16, %rd10, %rd8;
|
|
ld.global.v4.f32 {%f6,%f7,%f8,%f9}, [%rd15+0];
|
|
st.shared.v4.f32 [%rd16+0], {%f6,%f7,%f8,%f9};
|
|
$Lt_1_25602:
|
|
mov.u64 %rd7, __cuda___cuda_local_var_32689_34_non_const_lj13360;
|
|
mov.u64 %rd8, __cuda___cuda_local_var_32690_34_non_const_lj35296;
|
|
.loc 16 323 0
|
|
mov.f32 %f10, 0f00000000; // 0
|
|
mov.f32 %f11, %f10;
|
|
mov.f32 %f12, 0f00000000; // 0
|
|
mov.f32 %f13, %f12;
|
|
mov.f32 %f14, 0f00000000; // 0
|
|
mov.f32 %f15, %f14;
|
|
mov.f32 %f16, 0f00000000; // 0
|
|
mov.f32 %f17, %f16;
|
|
mov.f32 %f18, 0f00000000; // 0
|
|
mov.f32 %f19, %f18;
|
|
mov.f32 %f20, 0f00000000; // 0
|
|
mov.f32 %f21, %f20;
|
|
.loc 16 325 0
|
|
bar.sync 0;
|
|
ld.param.s32 %r4, [__cudaparm_kernel_pair_fast_t_per_atom];
|
|
div.s32 %r5, %r1, %r4;
|
|
cvt.s32.u32 %r6, %ntid.x;
|
|
div.s32 %r7, %r6, %r4;
|
|
rem.s32 %r8, %r1, %r4;
|
|
cvt.s32.u32 %r9, %ctaid.x;
|
|
mul.lo.s32 %r10, %r9, %r7;
|
|
add.s32 %r11, %r5, %r10;
|
|
ld.param.s32 %r12, [__cudaparm_kernel_pair_fast_inum];
|
|
setp.lt.s32 %p3, %r11, %r12;
|
|
@!%p3 bra $Lt_1_26370;
|
|
.loc 16 329 0
|
|
cvt.s64.s32 %rd17, %r11;
|
|
mul.wide.s32 %rd18, %r11, 4;
|
|
ld.param.u64 %rd19, [__cudaparm_kernel_pair_fast_dev_nbor];
|
|
add.u64 %rd20, %rd18, %rd19;
|
|
ld.global.s32 %r13, [%rd20+0];
|
|
.loc 16 331 0
|
|
ld.param.s32 %r14, [__cudaparm_kernel_pair_fast_nbor_pitch];
|
|
cvt.s64.s32 %rd21, %r14;
|
|
mul.wide.s32 %rd22, %r14, 4;
|
|
add.u64 %rd23, %rd22, %rd20;
|
|
ld.global.s32 %r15, [%rd23+0];
|
|
add.u64 %rd24, %rd22, %rd23;
|
|
ld.param.u64 %rd25, [__cudaparm_kernel_pair_fast_dev_packed];
|
|
setp.ne.u64 %p4, %rd25, %rd19;
|
|
@%p4 bra $Lt_1_26882;
|
|
.loc 16 337 0
|
|
cvt.s32.s64 %r16, %rd21;
|
|
mul.lo.s32 %r17, %r16, %r15;
|
|
cvt.s64.s32 %rd26, %r17;
|
|
mul.wide.s32 %rd27, %r17, 4;
|
|
add.u64 %rd28, %rd24, %rd27;
|
|
.loc 16 338 0
|
|
mul.lo.s32 %r18, %r8, %r16;
|
|
cvt.s64.s32 %rd29, %r18;
|
|
mul.wide.s32 %rd30, %r18, 4;
|
|
add.u64 %rd31, %rd24, %rd30;
|
|
.loc 16 339 0
|
|
mul.lo.s32 %r19, %r16, %r4;
|
|
bra.uni $Lt_1_26626;
|
|
$Lt_1_26882:
|
|
.loc 16 341 0
|
|
ld.global.s32 %r20, [%rd24+0];
|
|
cvt.s64.s32 %rd32, %r20;
|
|
mul.wide.s32 %rd33, %r20, 4;
|
|
add.u64 %rd34, %rd25, %rd33;
|
|
.loc 16 342 0
|
|
cvt.s64.s32 %rd35, %r15;
|
|
mul.wide.s32 %rd36, %r15, 4;
|
|
add.u64 %rd28, %rd34, %rd36;
|
|
.loc 16 343 0
|
|
mov.s32 %r19, %r4;
|
|
.loc 16 344 0
|
|
cvt.s64.s32 %rd37, %r8;
|
|
mul.wide.s32 %rd38, %r8, 4;
|
|
add.u64 %rd31, %rd34, %rd38;
|
|
$Lt_1_26626:
|
|
.loc 16 347 0
|
|
mov.u32 %r21, %r13;
|
|
mov.s32 %r22, 0;
|
|
mov.u32 %r23, %r22;
|
|
mov.s32 %r24, 0;
|
|
mov.u32 %r25, %r24;
|
|
mov.s32 %r26, 0;
|
|
mov.u32 %r27, %r26;
|
|
tex.1d.v4.f32.s32 {%f22,%f23,%f24,%f25},[pos_tex,{%r21,%r23,%r25,%r27}];
|
|
mov.f32 %f26, %f22;
|
|
mov.f32 %f27, %f23;
|
|
mov.f32 %f28, %f24;
|
|
mov.f32 %f29, %f25;
|
|
.loc 16 348 0
|
|
mov.u32 %r28, %r13;
|
|
mov.s32 %r29, 0;
|
|
mov.u32 %r30, %r29;
|
|
mov.s32 %r31, 0;
|
|
mov.u32 %r32, %r31;
|
|
mov.s32 %r33, 0;
|
|
mov.u32 %r34, %r33;
|
|
tex.1d.v4.f32.s32 {%f30,%f31,%f32,%f33},[q_tex,{%r28,%r30,%r32,%r34}];
|
|
mov.f32 %f34, %f30;
|
|
setp.ge.u64 %p5, %rd31, %rd28;
|
|
@%p5 bra $Lt_1_37890;
|
|
cvt.rzi.ftz.s32.f32 %r35, %f29;
|
|
cvt.s64.s32 %rd39, %r19;
|
|
mul.lo.s32 %r36, %r35, 11;
|
|
cvt.rn.f32.s32 %f35, %r36;
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
mov.f32 %f39, 0f00000000; // 0
|
|
mov.f32 %f40, 0f00000000; // 0
|
|
$Lt_1_27650:
|
|
//<loop> Loop body line 348, nesting depth: 1, estimated iterations: unknown
|
|
.loc 16 353 0
|
|
ld.global.s32 %r37, [%rd31+0];
|
|
.loc 16 356 0
|
|
shr.s32 %r38, %r37, 30;
|
|
and.b32 %r39, %r38, 3;
|
|
cvt.s64.s32 %rd40, %r39;
|
|
mul.wide.s32 %rd41, %r39, 4;
|
|
add.u64 %rd42, %rd1, %rd41;
|
|
ld.shared.f32 %f41, [%rd42+0];
|
|
.loc 16 357 0
|
|
mov.f32 %f42, 0f3f800000; // 1
|
|
ld.shared.f32 %f43, [%rd42+16];
|
|
sub.ftz.f32 %f44, %f42, %f43;
|
|
.loc 16 360 0
|
|
and.b32 %r40, %r37, 1073741823;
|
|
mov.u32 %r41, %r40;
|
|
mov.s32 %r42, 0;
|
|
mov.u32 %r43, %r42;
|
|
mov.s32 %r44, 0;
|
|
mov.u32 %r45, %r44;
|
|
mov.s32 %r46, 0;
|
|
mov.u32 %r47, %r46;
|
|
tex.1d.v4.f32.s32 {%f45,%f46,%f47,%f48},[pos_tex,{%r41,%r43,%r45,%r47}];
|
|
mov.f32 %f49, %f45;
|
|
mov.f32 %f50, %f46;
|
|
mov.f32 %f51, %f47;
|
|
mov.f32 %f52, %f48;
|
|
sub.ftz.f32 %f53, %f27, %f50;
|
|
sub.ftz.f32 %f54, %f26, %f49;
|
|
sub.ftz.f32 %f55, %f28, %f51;
|
|
mul.ftz.f32 %f56, %f53, %f53;
|
|
fma.rn.ftz.f32 %f57, %f54, %f54, %f56;
|
|
fma.rn.ftz.f32 %f58, %f55, %f55, %f57;
|
|
add.ftz.f32 %f59, %f35, %f52;
|
|
cvt.rzi.ftz.s32.f32 %r48, %f59;
|
|
cvt.s64.s32 %rd43, %r48;
|
|
mul.wide.s32 %rd44, %r48, 16;
|
|
add.u64 %rd45, %rd44, %rd7;
|
|
ld.shared.f32 %f60, [%rd45+0];
|
|
setp.gt.ftz.f32 %p6, %f60, %f58;
|
|
@!%p6 bra $Lt_1_31490;
|
|
rcp.approx.ftz.f32 %f61, %f58;
|
|
ld.shared.f32 %f62, [%rd45+4];
|
|
setp.lt.ftz.f32 %p7, %f58, %f62;
|
|
@!%p7 bra $Lt_1_28674;
|
|
add.u64 %rd46, %rd44, %rd8;
|
|
ld.shared.f32 %f63, [%rd46+0];
|
|
mov.f32 %f64, 0f40000000; // 2
|
|
setp.eq.ftz.f32 %p8, %f63, %f64;
|
|
@!%p8 bra $Lt_1_29186;
|
|
.loc 16 375 0
|
|
mul.ftz.f32 %f65, %f61, %f61;
|
|
mov.f32 %f66, %f65;
|
|
mov.f32 %f67, %f66;
|
|
.loc 16 376 0
|
|
mul.ftz.f32 %f68, %f65, %f65;
|
|
mov.f32 %f69, %f68;
|
|
bra.uni $Lt_1_29442;
|
|
$Lt_1_29186:
|
|
mov.f32 %f70, 0f3f800000; // 1
|
|
setp.eq.ftz.f32 %p9, %f63, %f70;
|
|
@!%p9 bra $Lt_1_29698;
|
|
.loc 16 378 0
|
|
sqrt.approx.ftz.f32 %f71, %f61;
|
|
mul.ftz.f32 %f72, %f61, %f71;
|
|
mov.f32 %f68, %f72;
|
|
mov.f32 %f69, %f68;
|
|
.loc 16 379 0
|
|
mul.ftz.f32 %f66, %f72, %f72;
|
|
mov.f32 %f67, %f66;
|
|
bra.uni $Lt_1_29442;
|
|
$Lt_1_29698:
|
|
.loc 16 381 0
|
|
mul.ftz.f32 %f73, %f61, %f61;
|
|
mul.ftz.f32 %f74, %f61, %f73;
|
|
mov.f32 %f66, %f74;
|
|
mov.f32 %f67, %f66;
|
|
.loc 16 382 0
|
|
mov.f32 %f68, %f74;
|
|
mov.f32 %f69, %f68;
|
|
$Lt_1_29442:
|
|
$Lt_1_28930:
|
|
.loc 16 384 0
|
|
mul.ftz.f32 %f75, %f41, %f66;
|
|
ld.shared.v2.f32 {%f76,%f77}, [%rd45+8];
|
|
mul.ftz.f32 %f78, %f76, %f68;
|
|
sub.ftz.f32 %f79, %f78, %f77;
|
|
mul.ftz.f32 %f80, %f75, %f79;
|
|
bra.uni $Lt_1_28418;
|
|
$Lt_1_28674:
|
|
.loc 16 386 0
|
|
mov.f32 %f80, 0f00000000; // 0
|
|
$Lt_1_28418:
|
|
ld.param.f32 %f81, [__cudaparm_kernel_pair_fast_cut_coulsq];
|
|
setp.gt.ftz.f32 %p10, %f81, %f58;
|
|
@!%p10 bra $Lt_1_30210;
|
|
.loc 16 393 0
|
|
sqrt.approx.ftz.f32 %f82, %f58;
|
|
ld.param.f32 %f83, [__cudaparm_kernel_pair_fast_g_ewald];
|
|
mul.ftz.f32 %f84, %f83, %f82;
|
|
mul.ftz.f32 %f85, %f84, %f84;
|
|
mov.f32 %f86, 0f3f800000; // 1
|
|
mov.f32 %f87, 0f3ea7ba05; // 0.327591
|
|
fma.rn.ftz.f32 %f88, %f87, %f84, %f86;
|
|
neg.ftz.f32 %f89, %f85;
|
|
rcp.approx.ftz.f32 %f90, %f88;
|
|
mov.f32 %f91, 0f3fb8aa3b; // 1.4427
|
|
mul.ftz.f32 %f92, %f89, %f91;
|
|
ex2.approx.ftz.f32 %f93, %f92;
|
|
mov.f32 %f94, 0f3e827906; // 0.25483
|
|
mov.f32 %f95, 0fbe91a98e; // -0.284497
|
|
mov.f32 %f96, 0f3fb5f0e3; // 1.42141
|
|
mov.f32 %f97, 0fbfba00e3; // -1.45315
|
|
mov.f32 %f98, 0f3f87dc22; // 1.06141
|
|
fma.rn.ftz.f32 %f99, %f98, %f90, %f97;
|
|
fma.rn.ftz.f32 %f100, %f90, %f99, %f96;
|
|
fma.rn.ftz.f32 %f101, %f90, %f100, %f95;
|
|
fma.rn.ftz.f32 %f102, %f90, %f101, %f94;
|
|
mul.ftz.f32 %f103, %f90, %f102;
|
|
mul.ftz.f32 %f104, %f93, %f103;
|
|
mov.f32 %f105, %f104;
|
|
.loc 16 394 0
|
|
mov.u32 %r49, %r40;
|
|
mov.s32 %r50, 0;
|
|
mov.u32 %r51, %r50;
|
|
mov.s32 %r52, 0;
|
|
mov.u32 %r53, %r52;
|
|
mov.s32 %r54, 0;
|
|
mov.u32 %r55, %r54;
|
|
tex.1d.v4.f32.s32 {%f106,%f107,%f108,%f109},[q_tex,{%r49,%r51,%r53,%r55}];
|
|
mov.f32 %f110, %f106;
|
|
ld.param.f32 %f111, [__cudaparm_kernel_pair_fast_qqrd2e];
|
|
mul.ftz.f32 %f112, %f111, %f34;
|
|
mul.ftz.f32 %f113, %f112, %f110;
|
|
div.approx.ftz.f32 %f114, %f113, %f82;
|
|
mov.f32 %f115, %f114;
|
|
.loc 16 395 0
|
|
mov.f32 %f116, 0f3f906ebb; // 1.12838
|
|
mul.ftz.f32 %f117, %f84, %f116;
|
|
fma.rn.ftz.f32 %f118, %f93, %f117, %f104;
|
|
sub.ftz.f32 %f119, %f118, %f44;
|
|
mul.ftz.f32 %f120, %f114, %f119;
|
|
bra.uni $Lt_1_29954;
|
|
$Lt_1_30210:
|
|
.loc 16 397 0
|
|
mov.f32 %f120, 0f00000000; // 0
|
|
$Lt_1_29954:
|
|
.loc 16 401 0
|
|
add.ftz.f32 %f121, %f120, %f80;
|
|
mul.ftz.f32 %f122, %f121, %f61;
|
|
fma.rn.ftz.f32 %f38, %f54, %f122, %f38;
|
|
.loc 16 402 0
|
|
fma.rn.ftz.f32 %f37, %f53, %f122, %f37;
|
|
.loc 16 403 0
|
|
fma.rn.ftz.f32 %f36, %f55, %f122, %f36;
|
|
ld.param.s32 %r56, [__cudaparm_kernel_pair_fast_eflag];
|
|
mov.u32 %r57, 0;
|
|
setp.le.s32 %p11, %r56, %r57;
|
|
@%p11 bra $Lt_1_30978;
|
|
.loc 16 406 0
|
|
mov.f32 %f123, %f115;
|
|
mov.f32 %f124, %f105;
|
|
sub.ftz.f32 %f125, %f124, %f44;
|
|
fma.rn.ftz.f32 %f126, %f123, %f125, %f39;
|
|
selp.f32 %f39, %f126, %f39, %p10;
|
|
@!%p7 bra $Lt_1_30978;
|
|
.loc 16 409 0
|
|
add.u64 %rd47, %rd44, %rd8;
|
|
ld.shared.v4.f32 {_,%f127,%f128,%f129}, [%rd47+0];
|
|
mov.f32 %f130, %f67;
|
|
mul.ftz.f32 %f131, %f130, %f41;
|
|
mov.f32 %f132, %f69;
|
|
mul.ftz.f32 %f133, %f127, %f132;
|
|
sub.ftz.f32 %f134, %f133, %f128;
|
|
mul.ftz.f32 %f135, %f131, %f134;
|
|
sub.ftz.f32 %f136, %f135, %f129;
|
|
add.ftz.f32 %f40, %f40, %f136;
|
|
$Lt_1_30978:
|
|
$Lt_1_30466:
|
|
ld.param.s32 %r58, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r59, 0;
|
|
setp.le.s32 %p12, %r58, %r59;
|
|
@%p12 bra $Lt_1_31490;
|
|
.loc 16 414 0
|
|
mov.f32 %f137, %f11;
|
|
mul.ftz.f32 %f138, %f54, %f54;
|
|
fma.rn.ftz.f32 %f139, %f122, %f138, %f137;
|
|
mov.f32 %f11, %f139;
|
|
.loc 16 415 0
|
|
mov.f32 %f140, %f13;
|
|
fma.rn.ftz.f32 %f141, %f122, %f56, %f140;
|
|
mov.f32 %f13, %f141;
|
|
.loc 16 416 0
|
|
mov.f32 %f142, %f15;
|
|
mul.ftz.f32 %f143, %f55, %f55;
|
|
fma.rn.ftz.f32 %f144, %f122, %f143, %f142;
|
|
mov.f32 %f15, %f144;
|
|
.loc 16 417 0
|
|
mov.f32 %f145, %f17;
|
|
mul.ftz.f32 %f146, %f53, %f54;
|
|
fma.rn.ftz.f32 %f147, %f122, %f146, %f145;
|
|
mov.f32 %f17, %f147;
|
|
.loc 16 418 0
|
|
mov.f32 %f148, %f19;
|
|
mul.ftz.f32 %f149, %f54, %f55;
|
|
fma.rn.ftz.f32 %f150, %f122, %f149, %f148;
|
|
mov.f32 %f19, %f150;
|
|
.loc 16 419 0
|
|
mul.ftz.f32 %f151, %f53, %f55;
|
|
fma.rn.ftz.f32 %f20, %f122, %f151, %f20;
|
|
mov.f32 %f21, %f20;
|
|
$Lt_1_31490:
|
|
$Lt_1_27906:
|
|
.loc 16 352 0
|
|
mul.lo.u64 %rd48, %rd39, 4;
|
|
add.u64 %rd31, %rd31, %rd48;
|
|
setp.lt.u64 %p13, %rd31, %rd28;
|
|
@%p13 bra $Lt_1_27650;
|
|
bra.uni $Lt_1_26114;
|
|
$Lt_1_37890:
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
mov.f32 %f39, 0f00000000; // 0
|
|
mov.f32 %f40, 0f00000000; // 0
|
|
bra.uni $Lt_1_26114;
|
|
$Lt_1_26370:
|
|
mov.f32 %f36, 0f00000000; // 0
|
|
mov.f32 %f37, 0f00000000; // 0
|
|
mov.f32 %f38, 0f00000000; // 0
|
|
mov.f32 %f39, 0f00000000; // 0
|
|
mov.f32 %f40, 0f00000000; // 0
|
|
$Lt_1_26114:
|
|
mov.u32 %r60, 1;
|
|
setp.le.s32 %p14, %r4, %r60;
|
|
@%p14 bra $Lt_1_34306;
|
|
.loc 16 430 0
|
|
mov.u64 %rd49, __cuda___cuda_local_var_32812_35_non_const_red_acc7232;
|
|
cvt.s64.s32 %rd50, %r1;
|
|
mul.wide.s32 %rd51, %r1, 4;
|
|
add.u64 %rd52, %rd49, %rd51;
|
|
mov.f32 %f152, %f38;
|
|
st.shared.f32 [%rd52+0], %f152;
|
|
.loc 16 431 0
|
|
mov.f32 %f153, %f37;
|
|
st.shared.f32 [%rd52+512], %f153;
|
|
.loc 16 432 0
|
|
mov.f32 %f154, %f36;
|
|
st.shared.f32 [%rd52+1024], %f154;
|
|
.loc 16 433 0
|
|
mov.f32 %f155, %f40;
|
|
st.shared.f32 [%rd52+1536], %f155;
|
|
.loc 16 434 0
|
|
mov.f32 %f156, %f39;
|
|
st.shared.f32 [%rd52+2048], %f156;
|
|
.loc 16 436 0
|
|
shr.s32 %r61, %r4, 31;
|
|
mov.s32 %r62, 1;
|
|
and.b32 %r63, %r61, %r62;
|
|
add.s32 %r64, %r63, %r4;
|
|
shr.s32 %r65, %r64, 1;
|
|
mov.s32 %r66, %r65;
|
|
mov.u32 %r67, 0;
|
|
setp.ne.u32 %p15, %r65, %r67;
|
|
@!%p15 bra $Lt_1_32770;
|
|
$Lt_1_33282:
|
|
setp.ge.u32 %p16, %r8, %r66;
|
|
@%p16 bra $Lt_1_33538;
|
|
.loc 16 439 0
|
|
add.u32 %r68, %r1, %r66;
|
|
cvt.u64.u32 %rd53, %r68;
|
|
mul.wide.u32 %rd54, %r68, 4;
|
|
add.u64 %rd55, %rd49, %rd54;
|
|
ld.shared.f32 %f157, [%rd55+0];
|
|
add.ftz.f32 %f152, %f157, %f152;
|
|
st.shared.f32 [%rd52+0], %f152;
|
|
ld.shared.f32 %f158, [%rd55+512];
|
|
add.ftz.f32 %f153, %f158, %f153;
|
|
st.shared.f32 [%rd52+512], %f153;
|
|
ld.shared.f32 %f159, [%rd55+1024];
|
|
add.ftz.f32 %f154, %f159, %f154;
|
|
st.shared.f32 [%rd52+1024], %f154;
|
|
ld.shared.f32 %f160, [%rd55+1536];
|
|
add.ftz.f32 %f155, %f160, %f155;
|
|
st.shared.f32 [%rd52+1536], %f155;
|
|
ld.shared.f32 %f161, [%rd55+2048];
|
|
add.ftz.f32 %f156, %f161, %f156;
|
|
st.shared.f32 [%rd52+2048], %f156;
|
|
$Lt_1_33538:
|
|
.loc 16 436 0
|
|
shr.u32 %r66, %r66, 1;
|
|
mov.u32 %r69, 0;
|
|
setp.ne.u32 %p17, %r66, %r69;
|
|
@%p17 bra $Lt_1_33282;
|
|
$Lt_1_32770:
|
|
.loc 16 443 0
|
|
mov.f32 %f38, %f152;
|
|
.loc 16 444 0
|
|
mov.f32 %f37, %f153;
|
|
.loc 16 445 0
|
|
mov.f32 %f36, %f154;
|
|
.loc 16 446 0
|
|
mov.f32 %f40, %f155;
|
|
.loc 16 447 0
|
|
mov.f32 %f39, %f156;
|
|
ld.param.s32 %r70, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r71, 0;
|
|
setp.le.s32 %p18, %r70, %r71;
|
|
@%p18 bra $Lt_1_34306;
|
|
.loc 16 451 0
|
|
mov.f32 %f152, %f11;
|
|
st.shared.f32 [%rd52+0], %f152;
|
|
mov.f32 %f153, %f13;
|
|
st.shared.f32 [%rd52+512], %f153;
|
|
mov.f32 %f154, %f15;
|
|
st.shared.f32 [%rd52+1024], %f154;
|
|
mov.f32 %f155, %f17;
|
|
st.shared.f32 [%rd52+1536], %f155;
|
|
mov.f32 %f156, %f19;
|
|
st.shared.f32 [%rd52+2048], %f156;
|
|
mov.f32 %f162, %f21;
|
|
st.shared.f32 [%rd52+2560], %f162;
|
|
.loc 16 453 0
|
|
mov.s32 %r72, %r65;
|
|
@!%p15 bra $Lt_1_34818;
|
|
$Lt_1_35330:
|
|
setp.ge.u32 %p19, %r8, %r72;
|
|
@%p19 bra $Lt_1_35586;
|
|
.loc 16 456 0
|
|
add.u32 %r73, %r1, %r72;
|
|
cvt.u64.u32 %rd56, %r73;
|
|
mul.wide.u32 %rd57, %r73, 4;
|
|
add.u64 %rd58, %rd49, %rd57;
|
|
ld.shared.f32 %f163, [%rd58+0];
|
|
add.ftz.f32 %f152, %f163, %f152;
|
|
st.shared.f32 [%rd52+0], %f152;
|
|
ld.shared.f32 %f164, [%rd58+512];
|
|
add.ftz.f32 %f153, %f164, %f153;
|
|
st.shared.f32 [%rd52+512], %f153;
|
|
ld.shared.f32 %f165, [%rd58+1024];
|
|
add.ftz.f32 %f154, %f165, %f154;
|
|
st.shared.f32 [%rd52+1024], %f154;
|
|
ld.shared.f32 %f166, [%rd58+1536];
|
|
add.ftz.f32 %f155, %f166, %f155;
|
|
st.shared.f32 [%rd52+1536], %f155;
|
|
ld.shared.f32 %f167, [%rd58+2048];
|
|
add.ftz.f32 %f156, %f167, %f156;
|
|
st.shared.f32 [%rd52+2048], %f156;
|
|
ld.shared.f32 %f168, [%rd58+2560];
|
|
add.ftz.f32 %f162, %f168, %f162;
|
|
st.shared.f32 [%rd52+2560], %f162;
|
|
$Lt_1_35586:
|
|
.loc 16 453 0
|
|
shr.u32 %r72, %r72, 1;
|
|
mov.u32 %r74, 0;
|
|
setp.ne.u32 %p20, %r72, %r74;
|
|
@%p20 bra $Lt_1_35330;
|
|
$Lt_1_34818:
|
|
.loc 16 461 0
|
|
mov.f32 %f11, %f152;
|
|
mov.f32 %f13, %f153;
|
|
mov.f32 %f15, %f154;
|
|
mov.f32 %f17, %f155;
|
|
mov.f32 %f19, %f156;
|
|
mov.f32 %f21, %f162;
|
|
$Lt_1_34306:
|
|
$Lt_1_32258:
|
|
selp.s32 %r75, 1, 0, %p3;
|
|
mov.s32 %r76, 0;
|
|
set.eq.u32.s32 %r77, %r8, %r76;
|
|
neg.s32 %r78, %r77;
|
|
and.b32 %r79, %r75, %r78;
|
|
mov.u32 %r80, 0;
|
|
setp.eq.s32 %p21, %r79, %r80;
|
|
@%p21 bra $Lt_1_36354;
|
|
.loc 16 467 0
|
|
cvt.s64.s32 %rd59, %r11;
|
|
ld.param.u64 %rd60, [__cudaparm_kernel_pair_fast_engv];
|
|
mul.wide.s32 %rd61, %r11, 4;
|
|
add.u64 %rd62, %rd60, %rd61;
|
|
ld.param.s32 %r81, [__cudaparm_kernel_pair_fast_eflag];
|
|
mov.u32 %r82, 0;
|
|
setp.le.s32 %p22, %r81, %r82;
|
|
@%p22 bra $Lt_1_36866;
|
|
.loc 16 469 0
|
|
st.global.f32 [%rd62+0], %f40;
|
|
.loc 16 470 0
|
|
cvt.s64.s32 %rd63, %r12;
|
|
mul.wide.s32 %rd64, %r12, 4;
|
|
add.u64 %rd65, %rd64, %rd62;
|
|
.loc 16 471 0
|
|
st.global.f32 [%rd65+0], %f39;
|
|
.loc 16 472 0
|
|
add.u64 %rd62, %rd64, %rd65;
|
|
$Lt_1_36866:
|
|
ld.param.s32 %r83, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r84, 0;
|
|
setp.le.s32 %p23, %r83, %r84;
|
|
@%p23 bra $Lt_1_37378;
|
|
.loc 16 476 0
|
|
mov.f32 %f169, %f11;
|
|
st.global.f32 [%rd62+0], %f169;
|
|
.loc 16 477 0
|
|
cvt.s64.s32 %rd66, %r12;
|
|
mul.wide.s32 %rd67, %r12, 4;
|
|
add.u64 %rd68, %rd67, %rd62;
|
|
.loc 16 476 0
|
|
mov.f32 %f170, %f13;
|
|
st.global.f32 [%rd68+0], %f170;
|
|
.loc 16 477 0
|
|
add.u64 %rd69, %rd67, %rd68;
|
|
.loc 16 476 0
|
|
mov.f32 %f171, %f15;
|
|
st.global.f32 [%rd69+0], %f171;
|
|
.loc 16 477 0
|
|
add.u64 %rd70, %rd67, %rd69;
|
|
.loc 16 476 0
|
|
mov.f32 %f172, %f17;
|
|
st.global.f32 [%rd70+0], %f172;
|
|
.loc 16 477 0
|
|
add.u64 %rd62, %rd67, %rd70;
|
|
.loc 16 476 0
|
|
mov.f32 %f173, %f19;
|
|
st.global.f32 [%rd62+0], %f173;
|
|
mov.f32 %f174, %f21;
|
|
add.u64 %rd71, %rd67, %rd62;
|
|
st.global.f32 [%rd71+0], %f174;
|
|
$Lt_1_37378:
|
|
.loc 16 480 0
|
|
ld.param.u64 %rd72, [__cudaparm_kernel_pair_fast_ans];
|
|
mul.lo.u64 %rd73, %rd59, 16;
|
|
add.u64 %rd74, %rd72, %rd73;
|
|
mov.f32 %f175, %f176;
|
|
st.global.v4.f32 [%rd74+0], {%f38,%f37,%f36,%f175};
|
|
$Lt_1_36354:
|
|
.loc 16 482 0
|
|
exit;
|
|
$LDWend_kernel_pair_fast:
|
|
} // kernel_pair_fast
|
|
|