117 lines
4.4 KiB
Plaintext
117 lines
4.4 KiB
Plaintext
|
//
|
||
|
|
// Generated by NVIDIA NVVM Compiler
|
||
|
|
//
|
||
|
|
// Compiler Build ID: CL-23083092
|
||
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
||
|
|
// Based on LLVM 3.4svn
|
||
|
|
//
|
||
|
|
|
||
|
|
.version 6.1
|
||
|
|
.target sm_30
|
||
|
|
.address_size 64
|
||
|
|
|
||
|
|
// .globl _Z6oxMainv
|
||
|
|
.global .align 8 .b8 pixelID[8];
|
||
|
|
.global .align 8 .b8 resolution[8];
|
||
|
|
.global .align 1 .b8 image[1];
|
||
|
|
.global .align 1 .b8 input_buffer[1];
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.reg .b16 %rs<4>;
|
||
|
|
.reg .f32 %f<4>;
|
||
|
|
.reg .b32 %r<25>;
|
||
|
|
.reg .b64 %rd<27>;
|
||
|
|
|
||
|
|
|
||
|
|
ld.global.v2.u32 {%r9, %r10}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd3, %r9;
|
||
|
|
cvt.u64.u32 %rd4, %r10;
|
||
|
|
mov.u64 %rd25, image;
|
||
|
|
cvta.global.u64 %rd2, %rd25;
|
||
|
|
mov.u32 %r7, 2;
|
||
|
|
mov.u32 %r6, 8;
|
||
|
|
mov.u64 %rd24, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd1), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd3, %rd4, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs1, [%rd1];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1, %rs1;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd9, %r13;
|
||
|
|
cvt.u64.u32 %rd10, %r14;
|
||
|
|
// inline asm
|
||
|
|
call (%rd7), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd9, %rd10, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs2, [%rd7+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f2, %rs2;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r17, %r18}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd15, %r17;
|
||
|
|
cvt.u64.u32 %rd16, %r18;
|
||
|
|
// inline asm
|
||
|
|
call (%rd13), _rt_buffer_get_64, (%rd2, %r7, %r6, %rd15, %rd16, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs3, [%rd13+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f3, %rs3;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r21, %r22}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd21, %r21;
|
||
|
|
cvt.u64.u32 %rd22, %r22;
|
||
|
|
mov.u64 %rd26, input_buffer;
|
||
|
|
cvta.global.u64 %rd20, %rd26;
|
||
|
|
mov.u32 %r8, 12;
|
||
|
|
// inline asm
|
||
|
|
call (%rd19), _rt_buffer_get_64, (%rd20, %r7, %r8, %rd21, %rd22, %rd24, %rd24);
|
||
|
|
// inline asm
|
||
|
|
st.f32 [%rd19+8], %f3;
|
||
|
|
st.f32 [%rd19+4], %f2;
|
||
|
|
st.f32 [%rd19], %f1;
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|