aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added cuda examples

Browse Source
This commit is contained in:
Evghenii
2013-11-04 11:44:49 +01:00
parent cb6614da42
commit cb7cbec0d5
226 changed files with 284385 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

213
examples_cuda/mandelbrot_tasks3d/mandel_task_cu.ptx Normal file
View File

@@ -0,0 +1,213 @@
//
// Generated by NVIDIA NVVM Compiler
// Compiler built on Thu Jul 18 02:37:37 2013 (1374107857)
// Cuda compilation tools, release 5.5, V5.5.0
//
.version 3.2
.target sm_35
.address_size 64
.file 1 "/home/evghenii/soft/ispc-code/ispc/examples/mandelbrot_tasks3d/mandel_task_cu.cu", 1383122156, 1370
.file 2 "/usr/local/cuda-5.5/bin/..//include/cuda_device_runtime_api.h", 1375338991, 7655
.file 3 "/usr/local/cuda-5.5/bin/..//include/device_functions.h", 1375338991, 185228
.extern .func (.param .b32 func_retval0) vprintf
(
.param .b64 vprintf_param_0,
.param .b64 vprintf_param_1
)
;
.global .align 1 .b8 $str[26] = {118, 101, 99, 116, 111, 114, 73, 110, 100, 101, 120, 61, 32, 37, 100, 32, 32, 98, 105, 100, 61, 32, 37, 100, 10, 0};
.weak .func (.param .b32 func_retval0) cudaMalloc(
.param .b64 cudaMalloc_param_0,
.param .b64 cudaMalloc_param_1
)
{
.reg .s32 %r<2>;
mov.u32 %r1, 30;
st.param.b32 [func_retval0+0], %r1;
.loc 2 66 3
ret;
}
.weak .func (.param .b32 func_retval0) cudaFuncGetAttributes(
.param .b64 cudaFuncGetAttributes_param_0,
.param .b64 cudaFuncGetAttributes_param_1
)
{
.reg .s32 %r<2>;
mov.u32 %r1, 30;
st.param.b32 [func_retval0+0], %r1;
.loc 2 71 3
ret;
}
.visible .entry mandelbrot_scanline(
.param .f32 mandelbrot_scanline_param_0,
.param .f32 mandelbrot_scanline_param_1,
.param .f32 mandelbrot_scanline_param_2,
.param .f32 mandelbrot_scanline_param_3,
.param .u32 mandelbrot_scanline_param_4,
.param .u32 mandelbrot_scanline_param_5,
.param .u32 mandelbrot_scanline_param_6,
.param .u32 mandelbrot_scanline_param_7,
.param .u32 mandelbrot_scanline_param_8,
.param .u64 mandelbrot_scanline_param_9
)
{
.local .align 8 .b8 __local_depot2[8];
.reg .b64 %SP;
.reg .b64 %SPL;
.reg .pred %p<9>;
.reg .s32 %r<40>;
.reg .f32 %f<20>;
.reg .s64 %rd<8>;
mov.u64 %SPL, __local_depot2;
cvta.local.u64 %SP, %SPL;
ld.param.f32 %f9, [mandelbrot_scanline_param_0];
ld.param.f32 %f10, [mandelbrot_scanline_param_1];
ld.param.f32 %f11, [mandelbrot_scanline_param_2];
ld.param.f32 %f12, [mandelbrot_scanline_param_3];
ld.param.u32 %r14, [mandelbrot_scanline_param_4];
ld.param.u32 %r17, [mandelbrot_scanline_param_5];
ld.param.u32 %r15, [mandelbrot_scanline_param_6];
ld.param.u32 %r18, [mandelbrot_scanline_param_7];
ld.param.u32 %r16, [mandelbrot_scanline_param_8];
ld.param.u64 %rd1, [mandelbrot_scanline_param_9];
add.u64 %rd2, %SP, 0;
.loc 1 35 1
cvta.to.local.u64 %rd3, %rd2;
mov.u32 %r19, %tid.x;
and.b32 %r20, %r19, 31;
mov.u32 %r21, %ntid.x;
cvta.global.u64 %rd4, $str;
st.local.v2.u32 [%rd3], {%r20, %r21};
// Callseq Start 0
{
.reg .b32 temp_param_reg;
.param .b64 param0;
st.param.b64 [param0+0], %rd4;
.param .b64 param1;
st.param.b64 [param1+0], %rd2;
.param .b32 retval0;
.loc 1 35 1
call.uni (retval0),
vprintf,
(
param0,
param1
);
ld.param.b32 %r22, [retval0+0];
}
// Callseq End 0
.loc 1 36 1
mov.u32 %r23, %ctaid.x;
.loc 1 37 1
mad.lo.s32 %r24, %r23, %r15, %r15;
.loc 3 2621 10
min.s32 %r1, %r24, %r14;
.loc 1 39 1
mov.u32 %r25, %ctaid.y;
mul.lo.s32 %r37, %r25, %r18;
.loc 1 40 1
add.s32 %r26, %r37, %r18;
.loc 3 2621 10
min.s32 %r3, %r26, %r17;
.loc 1 42 1
setp.ge.s32 %p1, %r37, %r3;
@%p1 bra BB2_12;
cvta.to.global.u64 %rd5, %rd1;
BB2_2:
.loc 1 36 1
mul.lo.s32 %r38, %r23, %r15;
.loc 1 43 1
setp.ge.s32 %p2, %r38, %r1;
@%p2 bra BB2_11;
.loc 1 46 1
cvt.rn.f32.s32 %f13, %r37;
fma.rn.f32 %f1, %f13, %f12, %f11;
BB2_4:
.loc 1 45 1
add.s32 %r7, %r20, %r38;
cvt.rn.f32.u32 %f14, %r7;
fma.rn.f32 %f2, %f14, %f10, %f9;
mov.u32 %r39, 0;
setp.gt.s32 %p3, %r16, 0;
.loc 1 12 1
@%p3 bra BB2_5;
bra.uni BB2_8;
BB2_5:
mov.f32 %f18, %f1;
mov.f32 %f19, %f2;
BB2_6:
.loc 1 13 1
mov.f32 %f4, %f19;
mov.f32 %f3, %f18;
mul.f32 %f5, %f3, %f3;
mul.f32 %f6, %f4, %f4;
add.f32 %f15, %f6, %f5;
setp.gt.f32 %p4, %f15, 0f40800000;
@%p4 bra BB2_8;
.loc 1 16 1
sub.f32 %f16, %f6, %f5;
.loc 1 17 1
add.f32 %f17, %f4, %f4;
.loc 1 19 1
add.f32 %f7, %f2, %f16;
.loc 1 20 1
fma.rn.f32 %f8, %f17, %f3, %f1;
.loc 1 12 96
add.s32 %r39, %r39, 1;
.loc 1 12 1
setp.lt.s32 %p5, %r39, %r16;
mov.f32 %f18, %f8;
mov.f32 %f19, %f7;
@%p5 bra BB2_6;
BB2_8:
.loc 1 49 1
mad.lo.s32 %r34, %r37, %r14, %r38;
add.s32 %r11, %r34, %r20;
.loc 1 50 1
setp.ge.u32 %p6, %r7, %r1;
@%p6 bra BB2_10;
.loc 1 51 1
mul.wide.s32 %rd6, %r11, 4;
add.s64 %rd7, %rd5, %rd6;
st.global.u32 [%rd7], %r39;
BB2_10:
.loc 1 43 57
add.s32 %r38, %r38, 32;
.loc 1 43 1
setp.lt.s32 %p7, %r38, %r1;
@%p7 bra BB2_4;
BB2_11:
.loc 1 42 57
add.s32 %r37, %r37, 1;
.loc 1 42 1
setp.lt.s32 %p8, %r37, %r3;
@%p8 bra BB2_2;
BB2_12:
.loc 1 53 2
ret;
}