473 lines
15 KiB
Plaintext
473 lines
15 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 4 .b8 normal[12];
|
||
|
|
.global .align 4 .b8 camPos[12];
|
||
|
|
.global .align 4 .b8 root[4];
|
||
|
|
.global .align 4 .u32 imageEnabled;
|
||
|
|
.global .texref lightmap;
|
||
|
|
.global .align 16 .b8 tileInfo[16];
|
||
|
|
.global .align 4 .u32 additive;
|
||
|
|
.global .align 1 .b8 image[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 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 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 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 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 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.reg .pred %p<14>;
|
||
|
|
.reg .b16 %rs<39>;
|
||
|
|
.reg .f32 %f<98>;
|
||
|
|
.reg .b32 %r<125>;
|
||
|
|
.reg .b64 %rd<223>;
|
||
|
|
|
||
|
|
|
||
|
|
ld.global.v2.u32 {%r18, %r19}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd9, %r18;
|
||
|
|
cvt.u64.u32 %rd10, %r19;
|
||
|
|
mov.u64 %rd13, image;
|
||
|
|
cvta.global.u64 %rd8, %rd13;
|
||
|
|
mov.u32 %r16, 2;
|
||
|
|
mov.u32 %r17, 8;
|
||
|
|
mov.u64 %rd12, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd7), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd9, %rd10, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs3, [%rd7+6];
|
||
|
|
mov.f32 %f46, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs2, %f46;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
setp.gt.u16 %p1, %rs3, %rs2;
|
||
|
|
@%p1 bra BB0_17;
|
||
|
|
|
||
|
|
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd16, %r31;
|
||
|
|
cvt.u64.u32 %rd17, %r32;
|
||
|
|
// inline asm
|
||
|
|
call (%rd14), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd16, %rd17, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs4, [%rd14];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f80, %rs4;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r35, %r36}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd22, %r35;
|
||
|
|
cvt.u64.u32 %rd23, %r36;
|
||
|
|
// inline asm
|
||
|
|
call (%rd20), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd22, %rd23, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs5, [%rd20+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f81, %rs5;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r39, %r40}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd28, %r39;
|
||
|
|
cvt.u64.u32 %rd29, %r40;
|
||
|
|
// inline asm
|
||
|
|
call (%rd26), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd28, %rd29, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs6, [%rd26+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f82, %rs6;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
||
|
|
setp.eq.s32 %p2, %r43, 0;
|
||
|
|
add.s32 %r46, %r43, -1;
|
||
|
|
cvt.u64.u32 %rd39, %r46;
|
||
|
|
selp.b64 %rd34, 0, %rd39, %p2;
|
||
|
|
setp.eq.s32 %p3, %r44, 0;
|
||
|
|
add.s32 %r48, %r44, -1;
|
||
|
|
cvt.u64.u32 %rd40, %r48;
|
||
|
|
selp.b64 %rd35, 0, %rd40, %p3;
|
||
|
|
ld.global.v2.u32 {%r49, %r50}, [resolution];
|
||
|
|
add.s32 %r52, %r49, -1;
|
||
|
|
setp.eq.s32 %p4, %r43, %r52;
|
||
|
|
add.s32 %r53, %r43, 1;
|
||
|
|
selp.b32 %r1, %r52, %r53, %p4;
|
||
|
|
add.s32 %r55, %r50, -1;
|
||
|
|
setp.eq.s32 %p5, %r44, %r55;
|
||
|
|
add.s32 %r56, %r44, 1;
|
||
|
|
selp.b32 %r2, %r55, %r56, %p5;
|
||
|
|
// inline asm
|
||
|
|
call (%rd32), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs7, [%rd32+6];
|
||
|
|
mov.u32 %r119, 0;
|
||
|
|
setp.le.u16 %p6, %rs7, %rs2;
|
||
|
|
@%p6 bra BB0_3;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd41), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs8, [%rd41];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f50, %rs8;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd47), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs9, [%rd47+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f51, %rs9;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd53), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs10, [%rd53+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f52, %rs10;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f50;
|
||
|
|
add.f32 %f81, %f81, %f51;
|
||
|
|
add.f32 %f82, %f82, %f52;
|
||
|
|
mov.u32 %r119, 1;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.u32 %rd62, [pixelID];
|
||
|
|
// inline asm
|
||
|
|
call (%rd60), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs11, [%rd60+6];
|
||
|
|
setp.le.u16 %p7, %rs11, %rs2;
|
||
|
|
@%p7 bra BB0_5;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd67), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs12, [%rd67];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f53, %rs12;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd73), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs13, [%rd73+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f54, %rs13;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd79), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd62, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs14, [%rd79+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f55, %rs14;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f53;
|
||
|
|
add.f32 %f81, %f81, %f54;
|
||
|
|
add.f32 %f82, %f82, %f55;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_5:
|
||
|
|
cvt.u64.u32 %rd88, %r1;
|
||
|
|
// inline asm
|
||
|
|
call (%rd86), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs15, [%rd86+6];
|
||
|
|
setp.le.u16 %p8, %rs15, %rs2;
|
||
|
|
@%p8 bra BB0_7;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd93), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs16, [%rd93];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f56, %rs16;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd99), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs17, [%rd99+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f57, %rs17;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd105), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd35, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs18, [%rd105+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f58, %rs18;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f56;
|
||
|
|
add.f32 %f81, %f81, %f57;
|
||
|
|
add.f32 %f82, %f82, %f58;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_7:
|
||
|
|
ld.global.u32 %rd115, [pixelID+4];
|
||
|
|
// inline asm
|
||
|
|
call (%rd112), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs19, [%rd112+6];
|
||
|
|
setp.le.u16 %p9, %rs19, %rs2;
|
||
|
|
@%p9 bra BB0_9;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd119), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs20, [%rd119];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f59, %rs20;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd125), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs21, [%rd125+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f60, %rs21;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd131), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs22, [%rd131+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f61, %rs22;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f59;
|
||
|
|
add.f32 %f81, %f81, %f60;
|
||
|
|
add.f32 %f82, %f82, %f61;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_9:
|
||
|
|
// inline asm
|
||
|
|
call (%rd138), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs23, [%rd138+6];
|
||
|
|
setp.le.u16 %p10, %rs23, %rs2;
|
||
|
|
@%p10 bra BB0_11;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd145), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs24, [%rd145];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f62, %rs24;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd151), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs25, [%rd151+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f63, %rs25;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd157), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd115, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs26, [%rd157+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f64, %rs26;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f62;
|
||
|
|
add.f32 %f81, %f81, %f63;
|
||
|
|
add.f32 %f82, %f82, %f64;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_11:
|
||
|
|
cvt.u64.u32 %rd167, %r2;
|
||
|
|
// inline asm
|
||
|
|
call (%rd164), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs27, [%rd164+6];
|
||
|
|
setp.le.u16 %p11, %rs27, %rs2;
|
||
|
|
@%p11 bra BB0_13;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd171), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs28, [%rd171];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f65, %rs28;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd177), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs29, [%rd177+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f66, %rs29;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd183), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd34, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs30, [%rd183+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f67, %rs30;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f65;
|
||
|
|
add.f32 %f81, %f81, %f66;
|
||
|
|
add.f32 %f82, %f82, %f67;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_13:
|
||
|
|
// inline asm
|
||
|
|
call (%rd190), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs31, [%rd190+6];
|
||
|
|
setp.le.u16 %p12, %rs31, %rs2;
|
||
|
|
@%p12 bra BB0_15;
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
call (%rd197), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs32, [%rd197];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f68, %rs32;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd203), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs33, [%rd203+2];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f69, %rs33;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd209), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd88, %rd167, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
ld.u16 %rs34, [%rd209+4];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f70, %rs34;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f80, %f80, %f68;
|
||
|
|
add.f32 %f81, %f81, %f69;
|
||
|
|
add.f32 %f82, %f82, %f70;
|
||
|
|
add.s32 %r119, %r119, 1;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
setp.eq.s32 %p13, %r119, 0;
|
||
|
|
@%p13 bra BB0_17;
|
||
|
|
|
||
|
|
cvt.rn.f32.u32 %f75, %r119;
|
||
|
|
rcp.rn.f32 %f76, %f75;
|
||
|
|
mul.f32 %f71, %f80, %f76;
|
||
|
|
mul.f32 %f72, %f81, %f76;
|
||
|
|
mul.f32 %f73, %f82, %f76;
|
||
|
|
ld.global.v2.u32 {%r114, %r115}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd218, %r114;
|
||
|
|
cvt.u64.u32 %rd219, %r115;
|
||
|
|
// inline asm
|
||
|
|
call (%rd216), _rt_buffer_get_64, (%rd8, %r16, %r17, %rd218, %rd219, %rd12, %rd12);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs37, %f73;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs36, %f72;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs35, %f71;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f74, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs38, %f74;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd216], {%rs35, %rs36, %rs37, %rs38};
|
||
|
|
|
||
|
|
BB0_17:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|