119 lines
3.9 KiB
Plaintext
119 lines
3.9 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_0000bb58_00000000-9_pair_gpu_nbor_kernel.cpp3.i (/home/sjplimp/ccBI#.bBFvWV)
|
|
//-----------------------------------------------------------
|
|
|
|
//-----------------------------------------------------------
|
|
// 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_0000bb58_00000000-8_pair_gpu_nbor_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 "pair_gpu_nbor_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"
|
|
|
|
|
|
.entry kernel_unpack (
|
|
.param .u64 __cudaparm_kernel_unpack_dev_nbor,
|
|
.param .u64 __cudaparm_kernel_unpack_dev_ij,
|
|
.param .s32 __cudaparm_kernel_unpack_inum)
|
|
{
|
|
.reg .u32 %r<11>;
|
|
.reg .u64 %rd<27>;
|
|
.reg .pred %p<5>;
|
|
.loc 16 29 0
|
|
$LDWbegin_kernel_unpack:
|
|
mov.u32 %r1, %ctaid.x;
|
|
mov.u32 %r2, %ntid.x;
|
|
mul.lo.u32 %r3, %r1, %r2;
|
|
mov.u32 %r4, %tid.x;
|
|
add.u32 %r5, %r4, %r3;
|
|
ld.param.s32 %r6, [__cudaparm_kernel_unpack_inum];
|
|
setp.le.s32 %p1, %r6, %r5;
|
|
@%p1 bra $Lt_0_2050;
|
|
.loc 16 35 0
|
|
cvt.s64.s32 %rd1, %r6;
|
|
ld.param.u64 %rd2, [__cudaparm_kernel_unpack_dev_nbor];
|
|
cvt.s64.s32 %rd3, %r5;
|
|
add.u64 %rd4, %rd3, %rd1;
|
|
mul.lo.u64 %rd5, %rd4, 4;
|
|
add.u64 %rd6, %rd2, %rd5;
|
|
ld.global.s32 %r7, [%rd6+0];
|
|
.loc 16 36 0
|
|
mul.wide.s32 %rd7, %r6, 4;
|
|
add.u64 %rd8, %rd6, %rd7;
|
|
mov.s64 %rd9, %rd8;
|
|
.loc 16 37 0
|
|
ld.param.u64 %rd10, [__cudaparm_kernel_unpack_dev_ij];
|
|
ld.global.s32 %r8, [%rd8+0];
|
|
cvt.s64.s32 %rd11, %r8;
|
|
mul.wide.s32 %rd12, %r8, 4;
|
|
add.u64 %rd13, %rd10, %rd12;
|
|
.loc 16 38 0
|
|
cvt.s64.s32 %rd14, %r7;
|
|
mul.wide.s32 %rd15, %r7, 4;
|
|
add.u64 %rd16, %rd15, %rd13;
|
|
setp.le.u64 %p2, %rd16, %rd13;
|
|
@%p2 bra $Lt_0_2562;
|
|
add.u64 %rd17, %rd15, 3;
|
|
shr.s64 %rd18, %rd17, 63;
|
|
mov.s64 %rd19, 3;
|
|
and.b64 %rd20, %rd18, %rd19;
|
|
add.s64 %rd21, %rd20, %rd17;
|
|
shr.s64 %rd22, %rd21, 2;
|
|
mov.s64 %rd23, 1;
|
|
max.s64 %rd24, %rd22, %rd23;
|
|
mov.s64 %rd25, %rd24;
|
|
$Lt_0_3074:
|
|
//<loop> Loop body line 38, nesting depth: 1, estimated iterations: unknown
|
|
.loc 16 41 0
|
|
ld.global.s32 %r9, [%rd13+0];
|
|
st.global.s32 [%rd9+0], %r9;
|
|
.loc 16 42 0
|
|
add.u64 %rd9, %rd7, %rd9;
|
|
.loc 16 40 0
|
|
add.u64 %rd13, %rd13, 4;
|
|
setp.gt.u64 %p3, %rd16, %rd13;
|
|
@%p3 bra $Lt_0_3074;
|
|
$Lt_0_2562:
|
|
$Lt_0_2050:
|
|
.loc 16 45 0
|
|
exit;
|
|
$LDWend_kernel_unpack:
|
|
} // kernel_unpack
|
|
|