111 lines
4.0 KiB
Plaintext
111 lines
4.0 KiB
Plaintext
.version 1.4
|
|
.target sm_10, map_f64_to_f32
|
|
// compiled with /opt/cuda-5.0/open64/lib//be
|
|
// nvopencc 4.1 built on 2012-09-21
|
|
|
|
//-----------------------------------------------------------
|
|
// Compiling /tmp/tmpxft_00005451_00000000-9_Test06.cpp3.i (/tmp/ccBI#.yMSmOL)
|
|
//-----------------------------------------------------------
|
|
|
|
//-----------------------------------------------------------
|
|
// Options:
|
|
//-----------------------------------------------------------
|
|
// Target:ptx, ISA:sm_10, Endian:little, Pointer Size:32
|
|
// -O3 (Optimization level)
|
|
// -g0 (Debug level)
|
|
// -m2 (Report advisories)
|
|
//-----------------------------------------------------------
|
|
|
|
.file 1 "<command-line>"
|
|
.file 2 "/usr/include/stdc-predef.h"
|
|
.file 3 "/tmp/tmpxft_00005451_00000000-8_Test06.cudafe2.gpu"
|
|
.file 4 "/usr/lib/gcc/i686-linux-gnu/4.8/include/stddef.h"
|
|
.file 5 "/opt/cuda/bin/../include/crt/device_runtime.h"
|
|
.file 6 "/opt/cuda/bin/../include/host_defines.h"
|
|
.file 7 "/opt/cuda/bin/../include/builtin_types.h"
|
|
.file 8 "/opt/cuda-5.0/include/device_types.h"
|
|
.file 9 "/opt/cuda-5.0/include/driver_types.h"
|
|
.file 10 "/opt/cuda-5.0/include/surface_types.h"
|
|
.file 11 "/opt/cuda-5.0/include/texture_types.h"
|
|
.file 12 "/opt/cuda-5.0/include/vector_types.h"
|
|
.file 13 "/opt/cuda/bin/../include/device_launch_parameters.h"
|
|
.file 14 "/opt/cuda-5.0/include/crt/storage_class.h"
|
|
.file 15 "Test06.cu"
|
|
.file 16 "/opt/cuda/bin/../include/common_functions.h"
|
|
.file 17 "/opt/cuda-5.0/include/math_functions.h"
|
|
.file 18 "/opt/cuda-5.0/include/math_constants.h"
|
|
.file 19 "/opt/cuda-5.0/include/device_functions.h"
|
|
.file 20 "/opt/cuda-5.0/include/sm_11_atomic_functions.h"
|
|
.file 21 "/opt/cuda-5.0/include/sm_12_atomic_functions.h"
|
|
.file 22 "/opt/cuda-5.0/include/sm_13_double_functions.h"
|
|
.file 23 "/opt/cuda-5.0/include/sm_20_atomic_functions.h"
|
|
.file 24 "/opt/cuda-5.0/include/sm_35_atomic_functions.h"
|
|
.file 25 "/opt/cuda-5.0/include/sm_20_intrinsics.h"
|
|
.file 26 "/opt/cuda-5.0/include/sm_30_intrinsics.h"
|
|
.file 27 "/opt/cuda-5.0/include/sm_35_intrinsics.h"
|
|
.file 28 "/opt/cuda-5.0/include/surface_functions.h"
|
|
.file 29 "/opt/cuda-5.0/include/texture_fetch_functions.h"
|
|
.file 30 "/opt/cuda-5.0/include/texture_indirect_functions.h"
|
|
.file 31 "/opt/cuda-5.0/include/surface_indirect_functions.h"
|
|
.file 32 "/opt/cuda-5.0/include/math_functions_dbl_ptx1.h"
|
|
|
|
|
|
.entry addDoubleMatrices (
|
|
.param .u32 __cudaparm_addDoubleMatrices_a,
|
|
.param .u32 __cudaparm_addDoubleMatrices_b,
|
|
.param .u32 __cudaparm_addDoubleMatrices_c,
|
|
.param .s32 __cudaparm_addDoubleMatrices_rows,
|
|
.param .s32 __cudaparm_addDoubleMatrices_cols)
|
|
{
|
|
.reg .u16 %rh<6>;
|
|
.reg .u32 %r<30>;
|
|
.reg .f64 %fd<5>;
|
|
.reg .pred %p<3>;
|
|
.loc 15 46 0
|
|
$LDWbegin_addDoubleMatrices:
|
|
mov.u16 %rh1, %ctaid.y;
|
|
mov.u16 %rh2, %ntid.y;
|
|
mul.wide.u16 %r1, %rh1, %rh2;
|
|
mov.u16 %rh3, %ctaid.x;
|
|
mov.u16 %rh4, %ntid.x;
|
|
mul.wide.u16 %r2, %rh3, %rh4;
|
|
cvt.u32.u16 %r3, %tid.y;
|
|
add.u32 %r4, %r3, %r1;
|
|
cvt.u32.u16 %r5, %tid.x;
|
|
add.u32 %r6, %r5, %r2;
|
|
ld.param.s32 %r7, [__cudaparm_addDoubleMatrices_rows];
|
|
set.gt.u32.s32 %r8, %r7, %r4;
|
|
neg.s32 %r9, %r8;
|
|
ld.param.s32 %r10, [__cudaparm_addDoubleMatrices_cols];
|
|
set.gt.u32.s32 %r11, %r10, %r6;
|
|
neg.s32 %r12, %r11;
|
|
and.b32 %r13, %r9, %r12;
|
|
mov.u32 %r14, 0;
|
|
setp.eq.s32 %p1, %r13, %r14;
|
|
@%p1 bra $Lt_0_1282;
|
|
.loc 15 51 0
|
|
mul.lo.u32 %r15, %r4, 4;
|
|
mul.lo.u32 %r16, %r6, 8;
|
|
ld.param.u32 %r17, [__cudaparm_addDoubleMatrices_a];
|
|
add.u32 %r18, %r17, %r15;
|
|
ld.global.u32 %r19, [%r18+0];
|
|
add.u32 %r20, %r19, %r16;
|
|
ld.global.f64 %fd1, [%r20+0];
|
|
ld.param.u32 %r21, [__cudaparm_addDoubleMatrices_b];
|
|
add.u32 %r22, %r21, %r15;
|
|
ld.global.u32 %r23, [%r22+0];
|
|
add.u32 %r24, %r23, %r16;
|
|
ld.global.f64 %fd2, [%r24+0];
|
|
add.f64 %fd3, %fd1, %fd2;
|
|
ld.param.u32 %r25, [__cudaparm_addDoubleMatrices_c];
|
|
add.u32 %r26, %r25, %r15;
|
|
ld.global.u32 %r27, [%r26+0];
|
|
add.u32 %r28, %r27, %r16;
|
|
st.global.f64 [%r28+0], %fd3;
|
|
$Lt_0_1282:
|
|
.loc 15 52 0
|
|
exit;
|
|
$LDWend_addDoubleMatrices:
|
|
} // addDoubleMatrices
|
|
|