// // 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/test.cu", 1383046614, 1449 .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 .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 _Z19mandelbrot_scanlineffffiiiiiPi( .param .f32 _Z19mandelbrot_scanlineffffiiiiiPi_param_0, .param .f32 _Z19mandelbrot_scanlineffffiiiiiPi_param_1, .param .f32 _Z19mandelbrot_scanlineffffiiiiiPi_param_2, .param .f32 _Z19mandelbrot_scanlineffffiiiiiPi_param_3, .param .u32 _Z19mandelbrot_scanlineffffiiiiiPi_param_4, .param .u32 _Z19mandelbrot_scanlineffffiiiiiPi_param_5, .param .u32 _Z19mandelbrot_scanlineffffiiiiiPi_param_6, .param .u32 _Z19mandelbrot_scanlineffffiiiiiPi_param_7, .param .u32 _Z19mandelbrot_scanlineffffiiiiiPi_param_8, .param .u64 _Z19mandelbrot_scanlineffffiiiiiPi_param_9 ) { .reg .pred %p<9>; .reg .s32 %r<36>; .reg .f32 %f<20>; .reg .s64 %rd<5>; ld.param.f32 %f9, [_Z19mandelbrot_scanlineffffiiiiiPi_param_0]; ld.param.f32 %f10, [_Z19mandelbrot_scanlineffffiiiiiPi_param_1]; ld.param.f32 %f11, [_Z19mandelbrot_scanlineffffiiiiiPi_param_2]; ld.param.f32 %f12, [_Z19mandelbrot_scanlineffffiiiiiPi_param_3]; ld.param.u32 %r14, [_Z19mandelbrot_scanlineffffiiiiiPi_param_4]; ld.param.u32 %r17, [_Z19mandelbrot_scanlineffffiiiiiPi_param_5]; ld.param.u32 %r15, [_Z19mandelbrot_scanlineffffiiiiiPi_param_6]; ld.param.u32 %r18, [_Z19mandelbrot_scanlineffffiiiiiPi_param_7]; ld.param.u32 %r16, [_Z19mandelbrot_scanlineffffiiiiiPi_param_8]; ld.param.u64 %rd1, [_Z19mandelbrot_scanlineffffiiiiiPi_param_9]; .loc 1 43 1 mov.u32 %r19, %ctaid.x; .loc 1 44 1 mad.lo.s32 %r20, %r19, %r15, %r15; .loc 3 2621 10 min.s32 %r1, %r20, %r14; .loc 1 46 1 mov.u32 %r21, %ctaid.y; mul.lo.s32 %r33, %r21, %r18; .loc 1 47 1 add.s32 %r22, %r33, %r18; .loc 3 2621 10 min.s32 %r3, %r22, %r17; .loc 1 49 1 setp.ge.s32 %p1, %r33, %r3; @%p1 bra BB2_12; cvta.to.global.u64 %rd2, %rd1; BB2_2: .loc 1 43 1 mul.lo.s32 %r34, %r19, %r15; .loc 1 50 1 setp.ge.s32 %p2, %r34, %r1; @%p2 bra BB2_11; .loc 1 53 1 cvt.rn.f32.s32 %f13, %r33; fma.rn.f32 %f1, %f13, %f12, %f11; BB2_4: .loc 1 52 1 mov.u32 %r26, %tid.x; and.b32 %r27, %r26, 31; add.s32 %r7, %r27, %r34; cvt.rn.f32.u32 %f14, %r7; fma.rn.f32 %f2, %f14, %f10, %f9; mov.u32 %r35, 0; setp.gt.s32 %p3, %r16, 0; .loc 1 13 1 @%p3 bra BB2_5; bra.uni BB2_8; BB2_5: mov.f32 %f18, %f1; mov.f32 %f19, %f2; BB2_6: .loc 1 14 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 17 1 sub.f32 %f16, %f6, %f5; .loc 1 18 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 13 96 add.s32 %r35, %r35, 1; .loc 1 13 1 setp.lt.s32 %p5, %r35, %r16; mov.f32 %f18, %f8; mov.f32 %f19, %f7; @%p5 bra BB2_6; BB2_8: .loc 1 56 1 mad.lo.s32 %r30, %r33, %r14, %r34; add.s32 %r11, %r30, %r27; .loc 1 57 1 setp.ge.u32 %p6, %r7, %r1; @%p6 bra BB2_10; mul.wide.s32 %rd3, %r11, 4; add.s64 %rd4, %rd2, %rd3; .loc 1 58 1 st.global.u32 [%rd4], %r35; BB2_10: .loc 1 50 57 add.s32 %r34, %r34, 32; .loc 1 50 1 setp.lt.s32 %p7, %r34, %r1; @%p7 bra BB2_4; BB2_11: .loc 1 49 57 add.s32 %r33, %r33, 1; .loc 1 49 1 setp.lt.s32 %p8, %r33, %r3; @%p8 bra BB2_2; BB2_12: .loc 1 60 2 ret; }