1037 lines
30 KiB
Plaintext
1037 lines
30 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_0000c02c_00000000-9_cmm_cut_gpu_kernel.cpp3.i (/home/sjplimp/ccBI#.cW08s4)
|
|
//-----------------------------------------------------------
|
|
|
|
//-----------------------------------------------------------
|
|
// 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_0000c02c_00000000-8_cmm_cut_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 "cmm_cut_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;
|
|
|
|
.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 .s32 __cudaparm_kernel_pair_t_per_atom)
|
|
{
|
|
.reg .u32 %r<72>;
|
|
.reg .u64 %rd<62>;
|
|
.reg .f32 %f<111>;
|
|
.reg .pred %p<21>;
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32497_33_non_const_sp_lj92[16];
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32590_35_non_const_red_acc108[3072];
|
|
// __cuda_local_var_32504_10_non_const_f = 48
|
|
// __cuda_local_var_32508_9_non_const_virial = 16
|
|
.loc 16 88 0
|
|
$LDWbegin_kernel_pair:
|
|
.loc 16 95 0
|
|
ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];
|
|
ldu.global.f32 %f1, [%rd1+0];
|
|
.loc 16 96 0
|
|
ld.global.f32 %f2, [%rd1+4];
|
|
.loc 16 97 0
|
|
ld.global.f32 %f3, [%rd1+8];
|
|
.loc 16 98 0
|
|
ld.global.f32 %f4, [%rd1+12];
|
|
st.shared.v4.f32 [__cuda___cuda_local_var_32497_33_non_const_sp_lj92+0], {%f1,%f2,%f3,%f4};
|
|
.loc 16 107 0
|
|
mov.f32 %f5, 0f00000000; // 0
|
|
mov.f32 %f6, %f5;
|
|
mov.f32 %f7, 0f00000000; // 0
|
|
mov.f32 %f8, %f7;
|
|
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;
|
|
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_20738;
|
|
.loc 16 114 0
|
|
ld.param.s32 %r11, [__cudaparm_kernel_pair_nbor_pitch];
|
|
cvt.s64.s32 %rd2, %r11;
|
|
mul.wide.s32 %rd3, %r11, 4;
|
|
cvt.s64.s32 %rd4, %r9;
|
|
mul.wide.s32 %rd5, %r9, 4;
|
|
ld.param.u64 %rd6, [__cudaparm_kernel_pair_dev_nbor];
|
|
add.u64 %rd7, %rd5, %rd6;
|
|
add.u64 %rd8, %rd3, %rd7;
|
|
ld.global.s32 %r12, [%rd8+0];
|
|
add.u64 %rd9, %rd3, %rd8;
|
|
ld.param.u64 %rd10, [__cudaparm_kernel_pair_dev_packed];
|
|
setp.ne.u64 %p2, %rd10, %rd6;
|
|
@%p2 bra $Lt_0_21250;
|
|
.loc 16 120 0
|
|
cvt.s32.s64 %r13, %rd2;
|
|
mul.lo.s32 %r14, %r13, %r12;
|
|
cvt.s64.s32 %rd11, %r14;
|
|
mul.wide.s32 %rd12, %r14, 4;
|
|
add.u64 %rd13, %rd9, %rd12;
|
|
.loc 16 121 0
|
|
mul.lo.s32 %r15, %r6, %r13;
|
|
cvt.s64.s32 %rd14, %r15;
|
|
mul.wide.s32 %rd15, %r15, 4;
|
|
add.u64 %rd16, %rd9, %rd15;
|
|
.loc 16 122 0
|
|
mul.lo.s32 %r16, %r13, %r1;
|
|
bra.uni $Lt_0_20994;
|
|
$Lt_0_21250:
|
|
.loc 16 124 0
|
|
ld.global.s32 %r17, [%rd9+0];
|
|
cvt.s64.s32 %rd17, %r17;
|
|
mul.wide.s32 %rd18, %r17, 4;
|
|
add.u64 %rd19, %rd10, %rd18;
|
|
.loc 16 125 0
|
|
cvt.s64.s32 %rd20, %r12;
|
|
mul.wide.s32 %rd21, %r12, 4;
|
|
add.u64 %rd13, %rd19, %rd21;
|
|
.loc 16 126 0
|
|
mov.s32 %r16, %r1;
|
|
.loc 16 127 0
|
|
cvt.s64.s32 %rd22, %r6;
|
|
mul.wide.s32 %rd23, %r6, 4;
|
|
add.u64 %rd16, %rd19, %rd23;
|
|
$Lt_0_20994:
|
|
.loc 16 130 0
|
|
ld.global.s32 %r18, [%rd7+0];
|
|
mov.u32 %r19, %r18;
|
|
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 {%f17,%f18,%f19,%f20},[pos_tex,{%r19,%r21,%r23,%r25}];
|
|
mov.f32 %f21, %f17;
|
|
mov.f32 %f22, %f18;
|
|
mov.f32 %f23, %f19;
|
|
mov.f32 %f24, %f20;
|
|
setp.ge.u64 %p3, %rd16, %rd13;
|
|
@%p3 bra $Lt_0_30722;
|
|
cvt.rzi.ftz.s32.f32 %r26, %f24;
|
|
cvt.s64.s32 %rd24, %r16;
|
|
ld.param.s32 %r27, [__cudaparm_kernel_pair_lj_types];
|
|
mul.lo.s32 %r28, %r27, %r26;
|
|
ld.param.u64 %rd25, [__cudaparm_kernel_pair_lj1];
|
|
mov.f32 %f25, 0f00000000; // 0
|
|
mov.f32 %f26, 0f00000000; // 0
|
|
mov.f32 %f27, 0f00000000; // 0
|
|
mov.f32 %f28, 0f00000000; // 0
|
|
mov.u64 %rd26, __cuda___cuda_local_var_32497_33_non_const_sp_lj92;
|
|
$Lt_0_22018:
|
|
//<loop> Loop body line 130, nesting depth: 1, estimated iterations: unknown
|
|
.loc 16 136 0
|
|
ld.global.s32 %r29, [%rd16+0];
|
|
.loc 16 137 0
|
|
shr.s32 %r30, %r29, 30;
|
|
and.b32 %r31, %r30, 3;
|
|
cvt.s64.s32 %rd27, %r31;
|
|
mul.wide.s32 %rd28, %r31, 4;
|
|
add.u64 %rd29, %rd26, %rd28;
|
|
ld.shared.f32 %f29, [%rd29+0];
|
|
.loc 16 140 0
|
|
and.b32 %r32, %r29, 1073741823;
|
|
mov.u32 %r33, %r32;
|
|
mov.s32 %r34, 0;
|
|
mov.u32 %r35, %r34;
|
|
mov.s32 %r36, 0;
|
|
mov.u32 %r37, %r36;
|
|
mov.s32 %r38, 0;
|
|
mov.u32 %r39, %r38;
|
|
tex.1d.v4.f32.s32 {%f30,%f31,%f32,%f33},[pos_tex,{%r33,%r35,%r37,%r39}];
|
|
mov.f32 %f34, %f30;
|
|
mov.f32 %f35, %f31;
|
|
mov.f32 %f36, %f32;
|
|
mov.f32 %f37, %f33;
|
|
cvt.rzi.ftz.s32.f32 %r40, %f37;
|
|
sub.ftz.f32 %f38, %f22, %f35;
|
|
sub.ftz.f32 %f39, %f21, %f34;
|
|
sub.ftz.f32 %f40, %f23, %f36;
|
|
mul.ftz.f32 %f41, %f38, %f38;
|
|
fma.rn.ftz.f32 %f42, %f39, %f39, %f41;
|
|
fma.rn.ftz.f32 %f43, %f40, %f40, %f42;
|
|
add.s32 %r41, %r40, %r28;
|
|
cvt.s64.s32 %rd30, %r41;
|
|
mul.wide.s32 %rd31, %r41, 16;
|
|
add.u64 %rd32, %rd31, %rd25;
|
|
ld.global.f32 %f44, [%rd32+0];
|
|
setp.gt.ftz.f32 %p4, %f44, %f43;
|
|
@!%p4 bra $Lt_0_24322;
|
|
rcp.approx.ftz.f32 %f45, %f43;
|
|
ld.global.f32 %f46, [%rd32+4];
|
|
mov.f32 %f47, 0f40000000; // 2
|
|
setp.eq.ftz.f32 %p5, %f46, %f47;
|
|
@!%p5 bra $Lt_0_23042;
|
|
.loc 16 155 0
|
|
mul.ftz.f32 %f48, %f45, %f45;
|
|
mov.f32 %f49, %f48;
|
|
.loc 16 156 0
|
|
mul.ftz.f32 %f50, %f48, %f48;
|
|
bra.uni $Lt_0_23298;
|
|
$Lt_0_23042:
|
|
mov.f32 %f51, 0f3f800000; // 1
|
|
setp.eq.ftz.f32 %p6, %f46, %f51;
|
|
@!%p6 bra $Lt_0_23554;
|
|
.loc 16 158 0
|
|
sqrt.approx.ftz.f32 %f52, %f45;
|
|
mul.ftz.f32 %f53, %f45, %f52;
|
|
mov.f32 %f50, %f53;
|
|
.loc 16 159 0
|
|
mul.ftz.f32 %f49, %f53, %f53;
|
|
bra.uni $Lt_0_23298;
|
|
$Lt_0_23554:
|
|
.loc 16 161 0
|
|
mul.ftz.f32 %f54, %f45, %f45;
|
|
mul.ftz.f32 %f55, %f45, %f54;
|
|
mov.f32 %f49, %f55;
|
|
.loc 16 162 0
|
|
mov.f32 %f50, %f55;
|
|
$Lt_0_23298:
|
|
$Lt_0_22786:
|
|
.loc 16 164 0
|
|
mul.ftz.f32 %f56, %f45, %f29;
|
|
mul.ftz.f32 %f57, %f49, %f56;
|
|
ld.global.v2.f32 {%f58,%f59}, [%rd32+8];
|
|
mul.ftz.f32 %f60, %f58, %f50;
|
|
sub.ftz.f32 %f61, %f60, %f59;
|
|
mul.ftz.f32 %f62, %f57, %f61;
|
|
.loc 16 166 0
|
|
fma.rn.ftz.f32 %f27, %f39, %f62, %f27;
|
|
.loc 16 167 0
|
|
fma.rn.ftz.f32 %f26, %f38, %f62, %f26;
|
|
.loc 16 168 0
|
|
fma.rn.ftz.f32 %f25, %f40, %f62, %f25;
|
|
ld.param.s32 %r42, [__cudaparm_kernel_pair_eflag];
|
|
mov.u32 %r43, 0;
|
|
setp.le.s32 %p7, %r42, %r43;
|
|
@%p7 bra $Lt_0_23810;
|
|
.loc 16 170 0
|
|
ld.param.u64 %rd33, [__cudaparm_kernel_pair_lj3];
|
|
add.u64 %rd34, %rd33, %rd31;
|
|
ld.global.v4.f32 {%f63,%f64,%f65,_}, [%rd34+0];
|
|
mul.ftz.f32 %f66, %f29, %f49;
|
|
mul.ftz.f32 %f67, %f63, %f50;
|
|
sub.ftz.f32 %f68, %f67, %f64;
|
|
mul.ftz.f32 %f69, %f66, %f68;
|
|
sub.ftz.f32 %f70, %f69, %f65;
|
|
add.ftz.f32 %f28, %f28, %f70;
|
|
$Lt_0_23810:
|
|
ld.param.s32 %r44, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r45, 0;
|
|
setp.le.s32 %p8, %r44, %r45;
|
|
@%p8 bra $Lt_0_24322;
|
|
.loc 16 173 0
|
|
mov.f32 %f71, %f6;
|
|
mul.ftz.f32 %f72, %f39, %f39;
|
|
fma.rn.ftz.f32 %f73, %f62, %f72, %f71;
|
|
mov.f32 %f6, %f73;
|
|
.loc 16 174 0
|
|
mov.f32 %f74, %f8;
|
|
fma.rn.ftz.f32 %f75, %f62, %f41, %f74;
|
|
mov.f32 %f8, %f75;
|
|
.loc 16 175 0
|
|
mov.f32 %f76, %f10;
|
|
mul.ftz.f32 %f77, %f40, %f40;
|
|
fma.rn.ftz.f32 %f78, %f62, %f77, %f76;
|
|
mov.f32 %f10, %f78;
|
|
.loc 16 176 0
|
|
mov.f32 %f79, %f12;
|
|
mul.ftz.f32 %f80, %f38, %f39;
|
|
fma.rn.ftz.f32 %f81, %f62, %f80, %f79;
|
|
mov.f32 %f12, %f81;
|
|
.loc 16 177 0
|
|
mov.f32 %f82, %f14;
|
|
mul.ftz.f32 %f83, %f39, %f40;
|
|
fma.rn.ftz.f32 %f84, %f62, %f83, %f82;
|
|
mov.f32 %f14, %f84;
|
|
.loc 16 178 0
|
|
mul.ftz.f32 %f85, %f38, %f40;
|
|
fma.rn.ftz.f32 %f15, %f62, %f85, %f15;
|
|
mov.f32 %f16, %f15;
|
|
$Lt_0_24322:
|
|
$Lt_0_22274:
|
|
.loc 16 134 0
|
|
mul.lo.u64 %rd35, %rd24, 4;
|
|
add.u64 %rd16, %rd16, %rd35;
|
|
setp.lt.u64 %p9, %rd16, %rd13;
|
|
@%p9 bra $Lt_0_22018;
|
|
bra.uni $Lt_0_20482;
|
|
$Lt_0_30722:
|
|
mov.f32 %f25, 0f00000000; // 0
|
|
mov.f32 %f26, 0f00000000; // 0
|
|
mov.f32 %f27, 0f00000000; // 0
|
|
mov.f32 %f28, 0f00000000; // 0
|
|
bra.uni $Lt_0_20482;
|
|
$Lt_0_20738:
|
|
mov.f32 %f25, 0f00000000; // 0
|
|
mov.f32 %f26, 0f00000000; // 0
|
|
mov.f32 %f27, 0f00000000; // 0
|
|
mov.f32 %f28, 0f00000000; // 0
|
|
$Lt_0_20482:
|
|
mov.u32 %r46, 1;
|
|
setp.le.s32 %p10, %r1, %r46;
|
|
@%p10 bra $Lt_0_27138;
|
|
.loc 16 189 0
|
|
mov.u64 %rd36, __cuda___cuda_local_var_32590_35_non_const_red_acc108;
|
|
cvt.s64.s32 %rd37, %r2;
|
|
mul.wide.s32 %rd38, %r2, 4;
|
|
add.u64 %rd39, %rd36, %rd38;
|
|
mov.f32 %f86, %f27;
|
|
st.shared.f32 [%rd39+0], %f86;
|
|
.loc 16 190 0
|
|
mov.f32 %f87, %f26;
|
|
st.shared.f32 [%rd39+512], %f87;
|
|
.loc 16 191 0
|
|
mov.f32 %f88, %f25;
|
|
st.shared.f32 [%rd39+1024], %f88;
|
|
.loc 16 192 0
|
|
mov.f32 %f89, %f28;
|
|
st.shared.f32 [%rd39+1536], %f89;
|
|
.loc 16 194 0
|
|
shr.s32 %r47, %r1, 31;
|
|
mov.s32 %r48, 1;
|
|
and.b32 %r49, %r47, %r48;
|
|
add.s32 %r50, %r49, %r1;
|
|
shr.s32 %r51, %r50, 1;
|
|
mov.s32 %r52, %r51;
|
|
mov.u32 %r53, 0;
|
|
setp.ne.u32 %p11, %r51, %r53;
|
|
@!%p11 bra $Lt_0_25602;
|
|
$Lt_0_26114:
|
|
setp.ge.u32 %p12, %r6, %r52;
|
|
@%p12 bra $Lt_0_26370;
|
|
.loc 16 197 0
|
|
add.u32 %r54, %r2, %r52;
|
|
cvt.u64.u32 %rd40, %r54;
|
|
mul.wide.u32 %rd41, %r54, 4;
|
|
add.u64 %rd42, %rd36, %rd41;
|
|
ld.shared.f32 %f90, [%rd42+0];
|
|
add.ftz.f32 %f86, %f90, %f86;
|
|
st.shared.f32 [%rd39+0], %f86;
|
|
ld.shared.f32 %f91, [%rd42+512];
|
|
add.ftz.f32 %f87, %f91, %f87;
|
|
st.shared.f32 [%rd39+512], %f87;
|
|
ld.shared.f32 %f92, [%rd42+1024];
|
|
add.ftz.f32 %f88, %f92, %f88;
|
|
st.shared.f32 [%rd39+1024], %f88;
|
|
ld.shared.f32 %f93, [%rd42+1536];
|
|
add.ftz.f32 %f89, %f93, %f89;
|
|
st.shared.f32 [%rd39+1536], %f89;
|
|
$Lt_0_26370:
|
|
.loc 16 194 0
|
|
shr.u32 %r52, %r52, 1;
|
|
mov.u32 %r55, 0;
|
|
setp.ne.u32 %p13, %r52, %r55;
|
|
@%p13 bra $Lt_0_26114;
|
|
$Lt_0_25602:
|
|
.loc 16 201 0
|
|
mov.f32 %f27, %f86;
|
|
.loc 16 202 0
|
|
mov.f32 %f26, %f87;
|
|
.loc 16 203 0
|
|
mov.f32 %f25, %f88;
|
|
.loc 16 204 0
|
|
mov.f32 %f28, %f89;
|
|
ld.param.s32 %r56, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r57, 0;
|
|
setp.le.s32 %p14, %r56, %r57;
|
|
@%p14 bra $Lt_0_27138;
|
|
.loc 16 208 0
|
|
mov.f32 %f86, %f6;
|
|
st.shared.f32 [%rd39+0], %f86;
|
|
mov.f32 %f87, %f8;
|
|
st.shared.f32 [%rd39+512], %f87;
|
|
mov.f32 %f88, %f10;
|
|
st.shared.f32 [%rd39+1024], %f88;
|
|
mov.f32 %f89, %f12;
|
|
st.shared.f32 [%rd39+1536], %f89;
|
|
mov.f32 %f94, %f14;
|
|
st.shared.f32 [%rd39+2048], %f94;
|
|
mov.f32 %f95, %f16;
|
|
st.shared.f32 [%rd39+2560], %f95;
|
|
.loc 16 210 0
|
|
mov.s32 %r58, %r51;
|
|
@!%p11 bra $Lt_0_27650;
|
|
$Lt_0_28162:
|
|
setp.ge.u32 %p15, %r6, %r58;
|
|
@%p15 bra $Lt_0_28418;
|
|
.loc 16 213 0
|
|
add.u32 %r59, %r2, %r58;
|
|
cvt.u64.u32 %rd43, %r59;
|
|
mul.wide.u32 %rd44, %r59, 4;
|
|
add.u64 %rd45, %rd36, %rd44;
|
|
ld.shared.f32 %f96, [%rd45+0];
|
|
add.ftz.f32 %f86, %f96, %f86;
|
|
st.shared.f32 [%rd39+0], %f86;
|
|
ld.shared.f32 %f97, [%rd45+512];
|
|
add.ftz.f32 %f87, %f97, %f87;
|
|
st.shared.f32 [%rd39+512], %f87;
|
|
ld.shared.f32 %f98, [%rd45+1024];
|
|
add.ftz.f32 %f88, %f98, %f88;
|
|
st.shared.f32 [%rd39+1024], %f88;
|
|
ld.shared.f32 %f99, [%rd45+1536];
|
|
add.ftz.f32 %f89, %f99, %f89;
|
|
st.shared.f32 [%rd39+1536], %f89;
|
|
ld.shared.f32 %f100, [%rd45+2048];
|
|
add.ftz.f32 %f94, %f100, %f94;
|
|
st.shared.f32 [%rd39+2048], %f94;
|
|
ld.shared.f32 %f101, [%rd45+2560];
|
|
add.ftz.f32 %f95, %f101, %f95;
|
|
st.shared.f32 [%rd39+2560], %f95;
|
|
$Lt_0_28418:
|
|
.loc 16 210 0
|
|
shr.u32 %r58, %r58, 1;
|
|
mov.u32 %r60, 0;
|
|
setp.ne.u32 %p16, %r58, %r60;
|
|
@%p16 bra $Lt_0_28162;
|
|
$Lt_0_27650:
|
|
.loc 16 218 0
|
|
mov.f32 %f6, %f86;
|
|
mov.f32 %f8, %f87;
|
|
mov.f32 %f10, %f88;
|
|
mov.f32 %f12, %f89;
|
|
mov.f32 %f14, %f94;
|
|
mov.f32 %f16, %f95;
|
|
$Lt_0_27138:
|
|
$Lt_0_25090:
|
|
selp.s32 %r61, 1, 0, %p1;
|
|
mov.s32 %r62, 0;
|
|
set.eq.u32.s32 %r63, %r6, %r62;
|
|
neg.s32 %r64, %r63;
|
|
and.b32 %r65, %r61, %r64;
|
|
mov.u32 %r66, 0;
|
|
setp.eq.s32 %p17, %r65, %r66;
|
|
@%p17 bra $Lt_0_29186;
|
|
.loc 16 224 0
|
|
cvt.s64.s32 %rd46, %r9;
|
|
ld.param.u64 %rd47, [__cudaparm_kernel_pair_engv];
|
|
mul.wide.s32 %rd48, %r9, 4;
|
|
add.u64 %rd49, %rd47, %rd48;
|
|
ld.param.s32 %r67, [__cudaparm_kernel_pair_eflag];
|
|
mov.u32 %r68, 0;
|
|
setp.le.s32 %p18, %r67, %r68;
|
|
@%p18 bra $Lt_0_29698;
|
|
.loc 16 226 0
|
|
st.global.f32 [%rd49+0], %f28;
|
|
.loc 16 227 0
|
|
cvt.s64.s32 %rd50, %r10;
|
|
mul.wide.s32 %rd51, %r10, 4;
|
|
add.u64 %rd49, %rd49, %rd51;
|
|
$Lt_0_29698:
|
|
ld.param.s32 %r69, [__cudaparm_kernel_pair_vflag];
|
|
mov.u32 %r70, 0;
|
|
setp.le.s32 %p19, %r69, %r70;
|
|
@%p19 bra $Lt_0_30210;
|
|
.loc 16 231 0
|
|
mov.f32 %f102, %f6;
|
|
st.global.f32 [%rd49+0], %f102;
|
|
.loc 16 232 0
|
|
cvt.s64.s32 %rd52, %r10;
|
|
mul.wide.s32 %rd53, %r10, 4;
|
|
add.u64 %rd54, %rd53, %rd49;
|
|
.loc 16 231 0
|
|
mov.f32 %f103, %f8;
|
|
st.global.f32 [%rd54+0], %f103;
|
|
.loc 16 232 0
|
|
add.u64 %rd55, %rd53, %rd54;
|
|
.loc 16 231 0
|
|
mov.f32 %f104, %f10;
|
|
st.global.f32 [%rd55+0], %f104;
|
|
.loc 16 232 0
|
|
add.u64 %rd56, %rd53, %rd55;
|
|
.loc 16 231 0
|
|
mov.f32 %f105, %f12;
|
|
st.global.f32 [%rd56+0], %f105;
|
|
.loc 16 232 0
|
|
add.u64 %rd49, %rd53, %rd56;
|
|
.loc 16 231 0
|
|
mov.f32 %f106, %f14;
|
|
st.global.f32 [%rd49+0], %f106;
|
|
mov.f32 %f107, %f16;
|
|
add.u64 %rd57, %rd53, %rd49;
|
|
st.global.f32 [%rd57+0], %f107;
|
|
$Lt_0_30210:
|
|
.loc 16 235 0
|
|
ld.param.u64 %rd58, [__cudaparm_kernel_pair_ans];
|
|
mul.lo.u64 %rd59, %rd46, 16;
|
|
add.u64 %rd60, %rd58, %rd59;
|
|
mov.f32 %f108, %f109;
|
|
st.global.v4.f32 [%rd60+0], {%f27,%f26,%f25,%f108};
|
|
$Lt_0_29186:
|
|
.loc 16 237 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 .s32 __cudaparm_kernel_pair_fast_t_per_atom)
|
|
{
|
|
.reg .u32 %r<74>;
|
|
.reg .u64 %rd<74>;
|
|
.reg .f32 %f<118>;
|
|
.reg .pred %p<24>;
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32656_33_non_const_sp_lj3268[16];
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32654_34_non_const_lj13296[1936];
|
|
.shared .align 16 .b8 __cuda___cuda_local_var_32655_34_non_const_lj35232[1936];
|
|
.shared .align 4 .b8 __cuda___cuda_local_var_32753_35_non_const_red_acc7168[3072];
|
|
// __cuda_local_var_32666_10_non_const_f = 48
|
|
// __cuda_local_var_32670_9_non_const_virial = 16
|
|
.loc 16 245 0
|
|
$LDWbegin_kernel_pair_fast:
|
|
cvt.s32.u32 %r1, %tid.x;
|
|
mov.u32 %r2, 3;
|
|
setp.gt.s32 %p1, %r1, %r2;
|
|
@%p1 bra $Lt_1_22786;
|
|
.loc 16 255 0
|
|
mov.u64 %rd1, __cuda___cuda_local_var_32656_33_non_const_sp_lj3268;
|
|
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_22786:
|
|
mov.u64 %rd1, __cuda___cuda_local_var_32656_33_non_const_sp_lj3268;
|
|
mov.u32 %r3, 120;
|
|
setp.gt.s32 %p2, %r1, %r3;
|
|
@%p2 bra $Lt_1_23298;
|
|
.loc 16 257 0
|
|
mov.u64 %rd7, __cuda___cuda_local_var_32654_34_non_const_lj13296;
|
|
cvt.s64.s32 %rd8, %r1;
|
|
mul.wide.s32 %rd9, %r1, 16;
|
|
ld.param.u64 %rd10, [__cudaparm_kernel_pair_fast_lj1_in];
|
|
add.u64 %rd11, %rd10, %rd9;
|
|
add.u64 %rd12, %rd9, %rd7;
|
|
ld.global.v4.f32 {%f2,%f3,%f4,%f5}, [%rd11+0];
|
|
st.shared.v4.f32 [%rd12+0], {%f2,%f3,%f4,%f5};
|
|
ld.param.s32 %r4, [__cudaparm_kernel_pair_fast_eflag];
|
|
mov.u32 %r5, 0;
|
|
setp.le.s32 %p3, %r4, %r5;
|
|
@%p3 bra $Lt_1_23810;
|
|
.loc 16 259 0
|
|
mov.u64 %rd13, __cuda___cuda_local_var_32655_34_non_const_lj35232;
|
|
ld.param.u64 %rd14, [__cudaparm_kernel_pair_fast_lj3_in];
|
|
add.u64 %rd15, %rd14, %rd9;
|
|
add.u64 %rd16, %rd9, %rd13;
|
|
ld.global.v4.f32 {%f6,%f7,%f8,%f9}, [%rd15+0];
|
|
st.shared.v4.f32 [%rd16+0], {%f6,%f7,%f8,%f9};
|
|
$Lt_1_23810:
|
|
mov.u64 %rd13, __cuda___cuda_local_var_32655_34_non_const_lj35232;
|
|
$Lt_1_23298:
|
|
mov.u64 %rd7, __cuda___cuda_local_var_32654_34_non_const_lj13296;
|
|
mov.u64 %rd13, __cuda___cuda_local_var_32655_34_non_const_lj35232;
|
|
.loc 16 269 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 16 271 0
|
|
bar.sync 0;
|
|
ld.param.s32 %r6, [__cudaparm_kernel_pair_fast_t_per_atom];
|
|
div.s32 %r7, %r1, %r6;
|
|
cvt.s32.u32 %r8, %ntid.x;
|
|
div.s32 %r9, %r8, %r6;
|
|
rem.s32 %r10, %r1, %r6;
|
|
cvt.s32.u32 %r11, %ctaid.x;
|
|
mul.lo.s32 %r12, %r11, %r9;
|
|
add.s32 %r13, %r7, %r12;
|
|
ld.param.s32 %r14, [__cudaparm_kernel_pair_fast_inum];
|
|
setp.lt.s32 %p4, %r13, %r14;
|
|
@!%p4 bra $Lt_1_24578;
|
|
.loc 16 277 0
|
|
ld.param.s32 %r15, [__cudaparm_kernel_pair_fast_nbor_pitch];
|
|
cvt.s64.s32 %rd17, %r15;
|
|
mul.wide.s32 %rd18, %r15, 4;
|
|
cvt.s64.s32 %rd19, %r13;
|
|
mul.wide.s32 %rd20, %r13, 4;
|
|
ld.param.u64 %rd21, [__cudaparm_kernel_pair_fast_dev_nbor];
|
|
add.u64 %rd22, %rd20, %rd21;
|
|
add.u64 %rd23, %rd18, %rd22;
|
|
ld.global.s32 %r16, [%rd23+0];
|
|
add.u64 %rd24, %rd18, %rd23;
|
|
ld.param.u64 %rd25, [__cudaparm_kernel_pair_fast_dev_packed];
|
|
setp.ne.u64 %p5, %rd25, %rd21;
|
|
@%p5 bra $Lt_1_25090;
|
|
.loc 16 283 0
|
|
cvt.s32.s64 %r17, %rd17;
|
|
mul.lo.s32 %r18, %r17, %r16;
|
|
cvt.s64.s32 %rd26, %r18;
|
|
mul.wide.s32 %rd27, %r18, 4;
|
|
add.u64 %rd28, %rd24, %rd27;
|
|
.loc 16 284 0
|
|
mul.lo.s32 %r19, %r10, %r17;
|
|
cvt.s64.s32 %rd29, %r19;
|
|
mul.wide.s32 %rd30, %r19, 4;
|
|
add.u64 %rd31, %rd24, %rd30;
|
|
.loc 16 285 0
|
|
mul.lo.s32 %r20, %r17, %r6;
|
|
bra.uni $Lt_1_24834;
|
|
$Lt_1_25090:
|
|
.loc 16 287 0
|
|
ld.global.s32 %r21, [%rd24+0];
|
|
cvt.s64.s32 %rd32, %r21;
|
|
mul.wide.s32 %rd33, %r21, 4;
|
|
add.u64 %rd34, %rd25, %rd33;
|
|
.loc 16 288 0
|
|
cvt.s64.s32 %rd35, %r16;
|
|
mul.wide.s32 %rd36, %r16, 4;
|
|
add.u64 %rd28, %rd34, %rd36;
|
|
.loc 16 289 0
|
|
mov.s32 %r20, %r6;
|
|
.loc 16 290 0
|
|
cvt.s64.s32 %rd37, %r10;
|
|
mul.wide.s32 %rd38, %r10, 4;
|
|
add.u64 %rd31, %rd34, %rd38;
|
|
$Lt_1_24834:
|
|
.loc 16 293 0
|
|
ld.global.s32 %r22, [%rd22+0];
|
|
mov.u32 %r23, %r22;
|
|
mov.s32 %r24, 0;
|
|
mov.u32 %r25, %r24;
|
|
mov.s32 %r26, 0;
|
|
mov.u32 %r27, %r26;
|
|
mov.s32 %r28, 0;
|
|
mov.u32 %r29, %r28;
|
|
tex.1d.v4.f32.s32 {%f22,%f23,%f24,%f25},[pos_tex,{%r23,%r25,%r27,%r29}];
|
|
mov.f32 %f26, %f22;
|
|
mov.f32 %f27, %f23;
|
|
mov.f32 %f28, %f24;
|
|
mov.f32 %f29, %f25;
|
|
setp.ge.u64 %p6, %rd31, %rd28;
|
|
@%p6 bra $Lt_1_34562;
|
|
cvt.rzi.ftz.s32.f32 %r30, %f29;
|
|
cvt.s64.s32 %rd39, %r20;
|
|
mul.lo.s32 %r31, %r30, 11;
|
|
cvt.rn.f32.s32 %f30, %r31;
|
|
mov.f32 %f31, 0f00000000; // 0
|
|
mov.f32 %f32, 0f00000000; // 0
|
|
mov.f32 %f33, 0f00000000; // 0
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
$Lt_1_25858:
|
|
//<loop> Loop body line 293, nesting depth: 1, estimated iterations: unknown
|
|
.loc 16 300 0
|
|
ld.global.s32 %r32, [%rd31+0];
|
|
.loc 16 301 0
|
|
shr.s32 %r33, %r32, 30;
|
|
and.b32 %r34, %r33, 3;
|
|
cvt.s64.s32 %rd40, %r34;
|
|
mul.wide.s32 %rd41, %r34, 4;
|
|
add.u64 %rd42, %rd1, %rd41;
|
|
ld.shared.f32 %f35, [%rd42+0];
|
|
.loc 16 304 0
|
|
and.b32 %r35, %r32, 1073741823;
|
|
mov.u32 %r36, %r35;
|
|
mov.s32 %r37, 0;
|
|
mov.u32 %r38, %r37;
|
|
mov.s32 %r39, 0;
|
|
mov.u32 %r40, %r39;
|
|
mov.s32 %r41, 0;
|
|
mov.u32 %r42, %r41;
|
|
tex.1d.v4.f32.s32 {%f36,%f37,%f38,%f39},[pos_tex,{%r36,%r38,%r40,%r42}];
|
|
mov.f32 %f40, %f36;
|
|
mov.f32 %f41, %f37;
|
|
mov.f32 %f42, %f38;
|
|
mov.f32 %f43, %f39;
|
|
sub.ftz.f32 %f44, %f27, %f41;
|
|
sub.ftz.f32 %f45, %f26, %f40;
|
|
sub.ftz.f32 %f46, %f28, %f42;
|
|
mul.ftz.f32 %f47, %f44, %f44;
|
|
fma.rn.ftz.f32 %f48, %f45, %f45, %f47;
|
|
fma.rn.ftz.f32 %f49, %f46, %f46, %f48;
|
|
add.ftz.f32 %f50, %f30, %f43;
|
|
cvt.rzi.ftz.s32.f32 %r43, %f50;
|
|
cvt.s64.s32 %rd43, %r43;
|
|
mul.wide.s32 %rd44, %r43, 16;
|
|
add.u64 %rd45, %rd44, %rd7;
|
|
ld.shared.f32 %f51, [%rd45+0];
|
|
setp.gt.ftz.f32 %p7, %f51, %f49;
|
|
@!%p7 bra $Lt_1_28162;
|
|
rcp.approx.ftz.f32 %f52, %f49;
|
|
ld.shared.f32 %f53, [%rd45+4];
|
|
mov.f32 %f54, 0f40000000; // 2
|
|
setp.eq.ftz.f32 %p8, %f53, %f54;
|
|
@!%p8 bra $Lt_1_26882;
|
|
.loc 16 318 0
|
|
mul.ftz.f32 %f55, %f52, %f52;
|
|
mov.f32 %f56, %f55;
|
|
.loc 16 319 0
|
|
mul.ftz.f32 %f57, %f55, %f55;
|
|
bra.uni $Lt_1_27138;
|
|
$Lt_1_26882:
|
|
mov.f32 %f58, 0f3f800000; // 1
|
|
setp.eq.ftz.f32 %p9, %f53, %f58;
|
|
@!%p9 bra $Lt_1_27394;
|
|
.loc 16 321 0
|
|
sqrt.approx.ftz.f32 %f59, %f52;
|
|
mul.ftz.f32 %f60, %f52, %f59;
|
|
mov.f32 %f57, %f60;
|
|
.loc 16 322 0
|
|
mul.ftz.f32 %f56, %f60, %f60;
|
|
bra.uni $Lt_1_27138;
|
|
$Lt_1_27394:
|
|
.loc 16 324 0
|
|
mul.ftz.f32 %f61, %f52, %f52;
|
|
mul.ftz.f32 %f62, %f52, %f61;
|
|
mov.f32 %f56, %f62;
|
|
.loc 16 325 0
|
|
mov.f32 %f57, %f62;
|
|
$Lt_1_27138:
|
|
$Lt_1_26626:
|
|
.loc 16 327 0
|
|
mul.ftz.f32 %f63, %f52, %f35;
|
|
mul.ftz.f32 %f64, %f56, %f63;
|
|
ld.shared.v2.f32 {%f65,%f66}, [%rd45+8];
|
|
mul.ftz.f32 %f67, %f65, %f57;
|
|
sub.ftz.f32 %f68, %f67, %f66;
|
|
mul.ftz.f32 %f69, %f64, %f68;
|
|
.loc 16 329 0
|
|
fma.rn.ftz.f32 %f33, %f45, %f69, %f33;
|
|
.loc 16 330 0
|
|
fma.rn.ftz.f32 %f32, %f44, %f69, %f32;
|
|
.loc 16 331 0
|
|
fma.rn.ftz.f32 %f31, %f46, %f69, %f31;
|
|
ld.param.s32 %r44, [__cudaparm_kernel_pair_fast_eflag];
|
|
mov.u32 %r45, 0;
|
|
setp.le.s32 %p10, %r44, %r45;
|
|
@%p10 bra $Lt_1_27650;
|
|
.loc 16 333 0
|
|
add.u64 %rd46, %rd44, %rd13;
|
|
ld.shared.v4.f32 {%f70,%f71,%f72,_}, [%rd46+0];
|
|
mul.ftz.f32 %f73, %f35, %f56;
|
|
mul.ftz.f32 %f74, %f70, %f57;
|
|
sub.ftz.f32 %f75, %f74, %f71;
|
|
mul.ftz.f32 %f76, %f73, %f75;
|
|
sub.ftz.f32 %f77, %f76, %f72;
|
|
add.ftz.f32 %f34, %f34, %f77;
|
|
$Lt_1_27650:
|
|
ld.param.s32 %r46, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r47, 0;
|
|
setp.le.s32 %p11, %r46, %r47;
|
|
@%p11 bra $Lt_1_28162;
|
|
.loc 16 336 0
|
|
mov.f32 %f78, %f11;
|
|
mul.ftz.f32 %f79, %f45, %f45;
|
|
fma.rn.ftz.f32 %f80, %f69, %f79, %f78;
|
|
mov.f32 %f11, %f80;
|
|
.loc 16 337 0
|
|
mov.f32 %f81, %f13;
|
|
fma.rn.ftz.f32 %f82, %f69, %f47, %f81;
|
|
mov.f32 %f13, %f82;
|
|
.loc 16 338 0
|
|
mov.f32 %f83, %f15;
|
|
mul.ftz.f32 %f84, %f46, %f46;
|
|
fma.rn.ftz.f32 %f85, %f69, %f84, %f83;
|
|
mov.f32 %f15, %f85;
|
|
.loc 16 339 0
|
|
mov.f32 %f86, %f17;
|
|
mul.ftz.f32 %f87, %f44, %f45;
|
|
fma.rn.ftz.f32 %f88, %f69, %f87, %f86;
|
|
mov.f32 %f17, %f88;
|
|
.loc 16 340 0
|
|
mov.f32 %f89, %f19;
|
|
mul.ftz.f32 %f90, %f45, %f46;
|
|
fma.rn.ftz.f32 %f91, %f69, %f90, %f89;
|
|
mov.f32 %f19, %f91;
|
|
.loc 16 341 0
|
|
mul.ftz.f32 %f92, %f44, %f46;
|
|
fma.rn.ftz.f32 %f20, %f69, %f92, %f20;
|
|
mov.f32 %f21, %f20;
|
|
$Lt_1_28162:
|
|
$Lt_1_26114:
|
|
.loc 16 298 0
|
|
mul.lo.u64 %rd47, %rd39, 4;
|
|
add.u64 %rd31, %rd31, %rd47;
|
|
setp.lt.u64 %p12, %rd31, %rd28;
|
|
@%p12 bra $Lt_1_25858;
|
|
bra.uni $Lt_1_24322;
|
|
$Lt_1_34562:
|
|
mov.f32 %f31, 0f00000000; // 0
|
|
mov.f32 %f32, 0f00000000; // 0
|
|
mov.f32 %f33, 0f00000000; // 0
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
bra.uni $Lt_1_24322;
|
|
$Lt_1_24578:
|
|
mov.f32 %f31, 0f00000000; // 0
|
|
mov.f32 %f32, 0f00000000; // 0
|
|
mov.f32 %f33, 0f00000000; // 0
|
|
mov.f32 %f34, 0f00000000; // 0
|
|
$Lt_1_24322:
|
|
mov.u32 %r48, 1;
|
|
setp.le.s32 %p13, %r6, %r48;
|
|
@%p13 bra $Lt_1_30978;
|
|
.loc 16 352 0
|
|
mov.u64 %rd48, __cuda___cuda_local_var_32753_35_non_const_red_acc7168;
|
|
cvt.s64.s32 %rd49, %r1;
|
|
mul.wide.s32 %rd50, %r1, 4;
|
|
add.u64 %rd51, %rd48, %rd50;
|
|
mov.f32 %f93, %f33;
|
|
st.shared.f32 [%rd51+0], %f93;
|
|
.loc 16 353 0
|
|
mov.f32 %f94, %f32;
|
|
st.shared.f32 [%rd51+512], %f94;
|
|
.loc 16 354 0
|
|
mov.f32 %f95, %f31;
|
|
st.shared.f32 [%rd51+1024], %f95;
|
|
.loc 16 355 0
|
|
mov.f32 %f96, %f34;
|
|
st.shared.f32 [%rd51+1536], %f96;
|
|
.loc 16 357 0
|
|
shr.s32 %r49, %r6, 31;
|
|
mov.s32 %r50, 1;
|
|
and.b32 %r51, %r49, %r50;
|
|
add.s32 %r52, %r51, %r6;
|
|
shr.s32 %r53, %r52, 1;
|
|
mov.s32 %r54, %r53;
|
|
mov.u32 %r55, 0;
|
|
setp.ne.u32 %p14, %r53, %r55;
|
|
@!%p14 bra $Lt_1_29442;
|
|
$Lt_1_29954:
|
|
setp.ge.u32 %p15, %r10, %r54;
|
|
@%p15 bra $Lt_1_30210;
|
|
.loc 16 360 0
|
|
add.u32 %r56, %r1, %r54;
|
|
cvt.u64.u32 %rd52, %r56;
|
|
mul.wide.u32 %rd53, %r56, 4;
|
|
add.u64 %rd54, %rd48, %rd53;
|
|
ld.shared.f32 %f97, [%rd54+0];
|
|
add.ftz.f32 %f93, %f97, %f93;
|
|
st.shared.f32 [%rd51+0], %f93;
|
|
ld.shared.f32 %f98, [%rd54+512];
|
|
add.ftz.f32 %f94, %f98, %f94;
|
|
st.shared.f32 [%rd51+512], %f94;
|
|
ld.shared.f32 %f99, [%rd54+1024];
|
|
add.ftz.f32 %f95, %f99, %f95;
|
|
st.shared.f32 [%rd51+1024], %f95;
|
|
ld.shared.f32 %f100, [%rd54+1536];
|
|
add.ftz.f32 %f96, %f100, %f96;
|
|
st.shared.f32 [%rd51+1536], %f96;
|
|
$Lt_1_30210:
|
|
.loc 16 357 0
|
|
shr.u32 %r54, %r54, 1;
|
|
mov.u32 %r57, 0;
|
|
setp.ne.u32 %p16, %r54, %r57;
|
|
@%p16 bra $Lt_1_29954;
|
|
$Lt_1_29442:
|
|
.loc 16 364 0
|
|
mov.f32 %f33, %f93;
|
|
.loc 16 365 0
|
|
mov.f32 %f32, %f94;
|
|
.loc 16 366 0
|
|
mov.f32 %f31, %f95;
|
|
.loc 16 367 0
|
|
mov.f32 %f34, %f96;
|
|
ld.param.s32 %r58, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r59, 0;
|
|
setp.le.s32 %p17, %r58, %r59;
|
|
@%p17 bra $Lt_1_30978;
|
|
.loc 16 371 0
|
|
mov.f32 %f93, %f11;
|
|
st.shared.f32 [%rd51+0], %f93;
|
|
mov.f32 %f94, %f13;
|
|
st.shared.f32 [%rd51+512], %f94;
|
|
mov.f32 %f95, %f15;
|
|
st.shared.f32 [%rd51+1024], %f95;
|
|
mov.f32 %f96, %f17;
|
|
st.shared.f32 [%rd51+1536], %f96;
|
|
mov.f32 %f101, %f19;
|
|
st.shared.f32 [%rd51+2048], %f101;
|
|
mov.f32 %f102, %f21;
|
|
st.shared.f32 [%rd51+2560], %f102;
|
|
.loc 16 373 0
|
|
mov.s32 %r60, %r53;
|
|
@!%p14 bra $Lt_1_31490;
|
|
$Lt_1_32002:
|
|
setp.ge.u32 %p18, %r10, %r60;
|
|
@%p18 bra $Lt_1_32258;
|
|
.loc 16 376 0
|
|
add.u32 %r61, %r1, %r60;
|
|
cvt.u64.u32 %rd55, %r61;
|
|
mul.wide.u32 %rd56, %r61, 4;
|
|
add.u64 %rd57, %rd48, %rd56;
|
|
ld.shared.f32 %f103, [%rd57+0];
|
|
add.ftz.f32 %f93, %f103, %f93;
|
|
st.shared.f32 [%rd51+0], %f93;
|
|
ld.shared.f32 %f104, [%rd57+512];
|
|
add.ftz.f32 %f94, %f104, %f94;
|
|
st.shared.f32 [%rd51+512], %f94;
|
|
ld.shared.f32 %f105, [%rd57+1024];
|
|
add.ftz.f32 %f95, %f105, %f95;
|
|
st.shared.f32 [%rd51+1024], %f95;
|
|
ld.shared.f32 %f106, [%rd57+1536];
|
|
add.ftz.f32 %f96, %f106, %f96;
|
|
st.shared.f32 [%rd51+1536], %f96;
|
|
ld.shared.f32 %f107, [%rd57+2048];
|
|
add.ftz.f32 %f101, %f107, %f101;
|
|
st.shared.f32 [%rd51+2048], %f101;
|
|
ld.shared.f32 %f108, [%rd57+2560];
|
|
add.ftz.f32 %f102, %f108, %f102;
|
|
st.shared.f32 [%rd51+2560], %f102;
|
|
$Lt_1_32258:
|
|
.loc 16 373 0
|
|
shr.u32 %r60, %r60, 1;
|
|
mov.u32 %r62, 0;
|
|
setp.ne.u32 %p19, %r60, %r62;
|
|
@%p19 bra $Lt_1_32002;
|
|
$Lt_1_31490:
|
|
.loc 16 381 0
|
|
mov.f32 %f11, %f93;
|
|
mov.f32 %f13, %f94;
|
|
mov.f32 %f15, %f95;
|
|
mov.f32 %f17, %f96;
|
|
mov.f32 %f19, %f101;
|
|
mov.f32 %f21, %f102;
|
|
$Lt_1_30978:
|
|
$Lt_1_28930:
|
|
selp.s32 %r63, 1, 0, %p4;
|
|
mov.s32 %r64, 0;
|
|
set.eq.u32.s32 %r65, %r10, %r64;
|
|
neg.s32 %r66, %r65;
|
|
and.b32 %r67, %r63, %r66;
|
|
mov.u32 %r68, 0;
|
|
setp.eq.s32 %p20, %r67, %r68;
|
|
@%p20 bra $Lt_1_33026;
|
|
.loc 16 387 0
|
|
cvt.s64.s32 %rd58, %r13;
|
|
ld.param.u64 %rd59, [__cudaparm_kernel_pair_fast_engv];
|
|
mul.wide.s32 %rd60, %r13, 4;
|
|
add.u64 %rd61, %rd59, %rd60;
|
|
ld.param.s32 %r69, [__cudaparm_kernel_pair_fast_eflag];
|
|
mov.u32 %r70, 0;
|
|
setp.le.s32 %p21, %r69, %r70;
|
|
@%p21 bra $Lt_1_33538;
|
|
.loc 16 389 0
|
|
st.global.f32 [%rd61+0], %f34;
|
|
.loc 16 390 0
|
|
cvt.s64.s32 %rd62, %r14;
|
|
mul.wide.s32 %rd63, %r14, 4;
|
|
add.u64 %rd61, %rd61, %rd63;
|
|
$Lt_1_33538:
|
|
ld.param.s32 %r71, [__cudaparm_kernel_pair_fast_vflag];
|
|
mov.u32 %r72, 0;
|
|
setp.le.s32 %p22, %r71, %r72;
|
|
@%p22 bra $Lt_1_34050;
|
|
.loc 16 394 0
|
|
mov.f32 %f109, %f11;
|
|
st.global.f32 [%rd61+0], %f109;
|
|
.loc 16 395 0
|
|
cvt.s64.s32 %rd64, %r14;
|
|
mul.wide.s32 %rd65, %r14, 4;
|
|
add.u64 %rd66, %rd65, %rd61;
|
|
.loc 16 394 0
|
|
mov.f32 %f110, %f13;
|
|
st.global.f32 [%rd66+0], %f110;
|
|
.loc 16 395 0
|
|
add.u64 %rd67, %rd65, %rd66;
|
|
.loc 16 394 0
|
|
mov.f32 %f111, %f15;
|
|
st.global.f32 [%rd67+0], %f111;
|
|
.loc 16 395 0
|
|
add.u64 %rd68, %rd65, %rd67;
|
|
.loc 16 394 0
|
|
mov.f32 %f112, %f17;
|
|
st.global.f32 [%rd68+0], %f112;
|
|
.loc 16 395 0
|
|
add.u64 %rd61, %rd65, %rd68;
|
|
.loc 16 394 0
|
|
mov.f32 %f113, %f19;
|
|
st.global.f32 [%rd61+0], %f113;
|
|
mov.f32 %f114, %f21;
|
|
add.u64 %rd69, %rd65, %rd61;
|
|
st.global.f32 [%rd69+0], %f114;
|
|
$Lt_1_34050:
|
|
.loc 16 398 0
|
|
ld.param.u64 %rd70, [__cudaparm_kernel_pair_fast_ans];
|
|
mul.lo.u64 %rd71, %rd58, 16;
|
|
add.u64 %rd72, %rd70, %rd71;
|
|
mov.f32 %f115, %f116;
|
|
st.global.v4.f32 [%rd72+0], {%f33,%f32,%f31,%f115};
|
|
$Lt_1_33026:
|
|
.loc 16 400 0
|
|
exit;
|
|
$LDWend_kernel_pair_fast:
|
|
} // kernel_pair_fast
|
|
|