383 lines
14 KiB
Plaintext
383 lines
14 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_HDR[1];
|
||
|
|
.global .align 1 .b8 image_HDR2[1];
|
||
|
|
.global .align 1 .b8 uvfacenormal[1];
|
||
|
|
.global .align 1 .b8 uvpos[1];
|
||
|
|
.global .align 1 .b8 uvsmoothpos[1];
|
||
|
|
.global .align 1 .b8 triangleMarks[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(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.local .align 4 .b8 __local_depot0[4];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<14>;
|
||
|
|
.reg .b16 %rs<10>;
|
||
|
|
.reg .f32 %f<160>;
|
||
|
|
.reg .b32 %r<86>;
|
||
|
|
.reg .b64 %rd<89>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd88, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd88;
|
||
|
|
ld.global.v2.u32 {%r6, %r7}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd4, %r6;
|
||
|
|
cvt.u64.u32 %rd5, %r7;
|
||
|
|
mov.u64 %rd8, uvfacenormal;
|
||
|
|
cvta.global.u64 %rd3, %rd8;
|
||
|
|
mov.u32 %r4, 2;
|
||
|
|
mov.u32 %r5, 4;
|
||
|
|
mov.u64 %rd7, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r4, %r5, %rd4, %rd5, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r1, [%rd2];
|
||
|
|
shr.u32 %r10, %r1, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r10;
|
||
|
|
and.b16 %rs3, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs4, %r1;
|
||
|
|
or.b16 %rs5, %rs4, %rs3;
|
||
|
|
setp.eq.s16 %p1, %rs5, 0;
|
||
|
|
mov.f32 %f150, 0f00000000;
|
||
|
|
mov.f32 %f151, %f150;
|
||
|
|
mov.f32 %f152, %f150;
|
||
|
|
@%p1 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs6, [%rd2+1];
|
||
|
|
and.b16 %rs8, %rs4, 255;
|
||
|
|
cvt.rn.f32.u16 %f61, %rs8;
|
||
|
|
div.rn.f32 %f62, %f61, 0f437F0000;
|
||
|
|
fma.rn.f32 %f63, %f62, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f64, %rs6;
|
||
|
|
div.rn.f32 %f65, %f64, 0f437F0000;
|
||
|
|
fma.rn.f32 %f66, %f65, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f67, %rs3;
|
||
|
|
div.rn.f32 %f68, %f67, 0f437F0000;
|
||
|
|
fma.rn.f32 %f69, %f68, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f70, %f66, %f66;
|
||
|
|
fma.rn.f32 %f71, %f63, %f63, %f70;
|
||
|
|
fma.rn.f32 %f72, %f69, %f69, %f71;
|
||
|
|
sqrt.rn.f32 %f73, %f72;
|
||
|
|
rcp.rn.f32 %f74, %f73;
|
||
|
|
mul.f32 %f150, %f63, %f74;
|
||
|
|
mul.f32 %f151, %f66, %f74;
|
||
|
|
mul.f32 %f152, %f69, %f74;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
ld.global.v2.u32 {%r11, %r12}, [pixelID];
|
||
|
|
ld.global.v2.u32 {%r14, %r15}, [tileInfo];
|
||
|
|
add.s32 %r2, %r11, %r14;
|
||
|
|
add.s32 %r3, %r12, %r15;
|
||
|
|
setp.eq.f32 %p2, %f151, 0f00000000;
|
||
|
|
setp.eq.f32 %p3, %f150, 0f00000000;
|
||
|
|
and.pred %p4, %p3, %p2;
|
||
|
|
setp.eq.f32 %p5, %f152, 0f00000000;
|
||
|
|
and.pred %p6, %p4, %p5;
|
||
|
|
@%p6 bra BB0_9;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_9:
|
||
|
|
cvt.u64.u32 %rd83, %r2;
|
||
|
|
cvt.u64.u32 %rd84, %r3;
|
||
|
|
mov.u64 %rd87, image_HDR;
|
||
|
|
cvta.global.u64 %rd82, %rd87;
|
||
|
|
mov.u32 %r85, 16;
|
||
|
|
// inline asm
|
||
|
|
call (%rd81), _rt_buffer_get_64, (%rd82, %r4, %r85, %rd83, %rd84, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f149, 0f00000000;
|
||
|
|
st.v4.f32 [%rd81], {%f149, %f149, %f149, %f149};
|
||
|
|
bra.uni BB0_10;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.v2.u32 {%r40, %r41}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd11, %r40;
|
||
|
|
cvt.u64.u32 %rd12, %r41;
|
||
|
|
mov.u64 %rd64, uvpos;
|
||
|
|
cvta.global.u64 %rd10, %rd64;
|
||
|
|
mov.u32 %r34, 16;
|
||
|
|
// inline asm
|
||
|
|
call (%rd9), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd11, %rd12, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f83, [%rd9];
|
||
|
|
ld.global.v2.u32 {%r44, %r45}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd17, %r44;
|
||
|
|
cvt.u64.u32 %rd18, %r45;
|
||
|
|
// inline asm
|
||
|
|
call (%rd15), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd17, %rd18, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f84, [%rd15+4];
|
||
|
|
ld.global.v2.u32 {%r48, %r49}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd23, %r48;
|
||
|
|
cvt.u64.u32 %rd24, %r49;
|
||
|
|
// inline asm
|
||
|
|
call (%rd21), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd23, %rd24, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f85, [%rd21+8];
|
||
|
|
ld.global.v2.u32 {%r52, %r53}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd29, %r52;
|
||
|
|
cvt.u64.u32 %rd30, %r53;
|
||
|
|
// inline asm
|
||
|
|
call (%rd27), _rt_buffer_get_64, (%rd10, %r4, %r34, %rd29, %rd30, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f86, [%rd27+12];
|
||
|
|
add.f32 %f82, %f86, 0f38D1B717;
|
||
|
|
ld.global.v2.u32 {%r56, %r57}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd35, %r56;
|
||
|
|
cvt.u64.u32 %rd36, %r57;
|
||
|
|
mov.u64 %rd65, uvsmoothpos;
|
||
|
|
cvta.global.u64 %rd34, %rd65;
|
||
|
|
// inline asm
|
||
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd35, %rd36, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f87, [%rd33];
|
||
|
|
ld.global.v2.u32 {%r60, %r61}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd41, %r60;
|
||
|
|
cvt.u64.u32 %rd42, %r61;
|
||
|
|
// inline asm
|
||
|
|
call (%rd39), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd41, %rd42, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f88, [%rd39+4];
|
||
|
|
ld.global.v2.u32 {%r64, %r65}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd47, %r64;
|
||
|
|
cvt.u64.u32 %rd48, %r65;
|
||
|
|
// inline asm
|
||
|
|
call (%rd45), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd47, %rd48, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f89, [%rd45+8];
|
||
|
|
ld.global.v2.u32 {%r68, %r69}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd53, %r68;
|
||
|
|
cvt.u64.u32 %rd54, %r69;
|
||
|
|
// inline asm
|
||
|
|
call (%rd51), _rt_buffer_get_64, (%rd34, %r4, %r34, %rd53, %rd54, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f90, [%rd51+12];
|
||
|
|
cvt.rzi.u32.f32 %r72, %f90;
|
||
|
|
cvt.u64.u32 %rd59, %r72;
|
||
|
|
mov.u64 %rd66, triangleMarks;
|
||
|
|
cvta.global.u64 %rd58, %rd66;
|
||
|
|
mov.u32 %r36, 1;
|
||
|
|
// inline asm
|
||
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r36, %r36, %rd59, %rd7, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
ld.u8 %rs2, [%rd57];
|
||
|
|
setp.eq.s16 %p7, %rs2, 0;
|
||
|
|
selp.f32 %f157, %f87, %f83, %p7;
|
||
|
|
selp.f32 %f158, %f88, %f84, %p7;
|
||
|
|
selp.f32 %f159, %f89, %f85, %p7;
|
||
|
|
fma.rn.f32 %f11, %f150, 0f38D1B717, %f157;
|
||
|
|
fma.rn.f32 %f12, %f151, 0f38D1B717, %f158;
|
||
|
|
fma.rn.f32 %f13, %f152, 0f38D1B717, %f159;
|
||
|
|
abs.f32 %f91, %f152;
|
||
|
|
abs.f32 %f92, %f150;
|
||
|
|
setp.gt.f32 %p8, %f92, %f91;
|
||
|
|
neg.f32 %f93, %f151;
|
||
|
|
selp.f32 %f94, %f93, 0f00000000, %p8;
|
||
|
|
neg.f32 %f95, %f152;
|
||
|
|
selp.f32 %f96, %f150, %f95, %p8;
|
||
|
|
selp.f32 %f97, 0f00000000, %f151, %p8;
|
||
|
|
mul.f32 %f98, %f96, %f96;
|
||
|
|
fma.rn.f32 %f99, %f94, %f94, %f98;
|
||
|
|
fma.rn.f32 %f100, %f97, %f97, %f99;
|
||
|
|
sqrt.rn.f32 %f101, %f100;
|
||
|
|
rcp.rn.f32 %f102, %f101;
|
||
|
|
mul.f32 %f14, %f94, %f102;
|
||
|
|
mul.f32 %f15, %f96, %f102;
|
||
|
|
mul.f32 %f16, %f97, %f102;
|
||
|
|
mul.f32 %f103, %f152, %f15;
|
||
|
|
mul.f32 %f104, %f151, %f16;
|
||
|
|
sub.f32 %f17, %f103, %f104;
|
||
|
|
mul.f32 %f105, %f150, %f16;
|
||
|
|
mul.f32 %f106, %f152, %f14;
|
||
|
|
sub.f32 %f18, %f105, %f106;
|
||
|
|
mul.f32 %f107, %f151, %f14;
|
||
|
|
mul.f32 %f108, %f150, %f15;
|
||
|
|
sub.f32 %f19, %f107, %f108;
|
||
|
|
mov.f32 %f109, 0f3F8147AE;
|
||
|
|
sqrt.rn.f32 %f110, %f109;
|
||
|
|
rcp.rn.f32 %f20, %f110;
|
||
|
|
neg.f32 %f21, %f20;
|
||
|
|
mul.f32 %f22, %f20, 0f00000000;
|
||
|
|
mul.f32 %f111, %f20, 0f3DCCCCCD;
|
||
|
|
mul.f32 %f23, %f14, %f22;
|
||
|
|
mul.f32 %f24, %f15, %f22;
|
||
|
|
mul.f32 %f25, %f16, %f22;
|
||
|
|
fma.rn.f32 %f112, %f17, %f21, %f23;
|
||
|
|
fma.rn.f32 %f113, %f18, %f21, %f24;
|
||
|
|
fma.rn.f32 %f114, %f19, %f21, %f25;
|
||
|
|
mul.f32 %f26, %f150, %f111;
|
||
|
|
mul.f32 %f27, %f151, %f111;
|
||
|
|
mul.f32 %f28, %f152, %f111;
|
||
|
|
add.f32 %f156, %f112, %f26;
|
||
|
|
add.f32 %f155, %f113, %f27;
|
||
|
|
add.f32 %f154, %f114, %f28;
|
||
|
|
add.u64 %rd63, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd67, %rd63;
|
||
|
|
mov.u32 %r38, 0;
|
||
|
|
st.local.u32 [%rd67], %r38;
|
||
|
|
ld.global.u32 %r37, [root];
|
||
|
|
mov.f32 %f81, 0f38D1B717;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r37, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f153, [%rd67];
|
||
|
|
setp.gt.f32 %p9, %f153, 0f00000000;
|
||
|
|
@%p9 bra BB0_7;
|
||
|
|
|
||
|
|
fma.rn.f32 %f123, %f17, %f20, %f23;
|
||
|
|
fma.rn.f32 %f124, %f18, %f20, %f24;
|
||
|
|
fma.rn.f32 %f125, %f19, %f20, %f25;
|
||
|
|
add.f32 %f156, %f123, %f26;
|
||
|
|
add.f32 %f155, %f124, %f27;
|
||
|
|
add.f32 %f154, %f125, %f28;
|
||
|
|
st.local.u32 [%rd67], %r38;
|
||
|
|
ld.global.u32 %r73, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r73, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f153, [%rd67];
|
||
|
|
setp.gt.f32 %p10, %f153, 0f00000000;
|
||
|
|
@%p10 bra BB0_7;
|
||
|
|
|
||
|
|
mul.f32 %f37, %f17, %f22;
|
||
|
|
fma.rn.f32 %f134, %f14, %f21, %f37;
|
||
|
|
mul.f32 %f38, %f18, %f22;
|
||
|
|
fma.rn.f32 %f135, %f15, %f21, %f38;
|
||
|
|
mul.f32 %f39, %f19, %f22;
|
||
|
|
fma.rn.f32 %f136, %f16, %f21, %f39;
|
||
|
|
add.f32 %f156, %f134, %f26;
|
||
|
|
add.f32 %f155, %f135, %f27;
|
||
|
|
add.f32 %f154, %f136, %f28;
|
||
|
|
st.local.u32 [%rd67], %r38;
|
||
|
|
ld.global.u32 %r76, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r76, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f153, [%rd67];
|
||
|
|
setp.gt.f32 %p11, %f153, 0f00000000;
|
||
|
|
@%p11 bra BB0_7;
|
||
|
|
|
||
|
|
fma.rn.f32 %f145, %f14, %f20, %f37;
|
||
|
|
fma.rn.f32 %f146, %f15, %f20, %f38;
|
||
|
|
fma.rn.f32 %f147, %f16, %f20, %f39;
|
||
|
|
add.f32 %f156, %f145, %f26;
|
||
|
|
add.f32 %f155, %f146, %f27;
|
||
|
|
add.f32 %f154, %f147, %f28;
|
||
|
|
st.local.u32 [%rd67], %r38;
|
||
|
|
ld.global.u32 %r79, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r79, %f11, %f12, %f13, %f156, %f155, %f154, %r38, %f81, %f82, %rd63, %r5);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f153, [%rd67];
|
||
|
|
setp.leu.f32 %p12, %f153, 0f00000000;
|
||
|
|
@%p12 bra BB0_8;
|
||
|
|
|
||
|
|
BB0_7:
|
||
|
|
fma.rn.f32 %f157, %f153, %f156, %f11;
|
||
|
|
fma.rn.f32 %f158, %f153, %f155, %f12;
|
||
|
|
fma.rn.f32 %f159, %f153, %f154, %f13;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
cvt.u64.u32 %rd77, %r3;
|
||
|
|
cvt.u64.u32 %rd76, %r2;
|
||
|
|
mov.u64 %rd80, image_HDR;
|
||
|
|
cvta.global.u64 %rd75, %rd80;
|
||
|
|
// inline asm
|
||
|
|
call (%rd74), _rt_buffer_get_64, (%rd75, %r4, %r34, %rd76, %rd77, %rd7, %rd7);
|
||
|
|
// inline asm
|
||
|
|
selp.f32 %f148, 0f3F000000, 0f3F800000, %p7;
|
||
|
|
st.v4.f32 [%rd74], {%f157, %f158, %f159, %f148};
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|