1621 lines
50 KiB
Plaintext
1621 lines
50 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 uvnormal[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 4 .b8 voxelSize[12];
|
|
.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 4 .b8 _ZN21rti_internal_typeinfo9voxelSizeE[8] = {82, 97, 121, 0, 12, 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 8 .b8 _ZN21rti_internal_typename9voxelSizeE[7] = {102, 108, 111, 97, 116, 51, 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 4 .u32 _ZN21rti_internal_typeenum9voxelSizeE = 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 _ZN21rti_internal_semantic9voxelSizeE[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];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9voxelSizeE[1];
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[24];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<73>;
|
|
.reg .b16 %rs<9>;
|
|
.reg .f32 %f<1247>;
|
|
.reg .b32 %r<137>;
|
|
.reg .b64 %rd<94>;
|
|
|
|
|
|
mov.u64 %rd93, __local_depot0;
|
|
cvta.local.u64 %SP, %rd93;
|
|
ld.global.v2.u32 {%r8, %r9}, [pixelID];
|
|
cvt.u64.u32 %rd6, %r8;
|
|
cvt.u64.u32 %rd7, %r9;
|
|
mov.u64 %rd10, uvnormal;
|
|
cvta.global.u64 %rd5, %rd10;
|
|
mov.u32 %r6, 2;
|
|
mov.u32 %r7, 4;
|
|
mov.u64 %rd9, 0;
|
|
// inline asm
|
|
call (%rd4), _rt_buffer_get_64, (%rd5, %r6, %r7, %rd6, %rd7, %rd9, %rd9);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd4];
|
|
shr.u32 %r12, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r12;
|
|
and.b16 %rs2, %rs1, 255;
|
|
cvt.u16.u32 %rs3, %r1;
|
|
or.b16 %rs4, %rs3, %rs2;
|
|
setp.eq.s16 %p10, %rs4, 0;
|
|
mov.pred %p70, -1;
|
|
mov.pred %p71, %p70;
|
|
mov.pred %p72, %p70;
|
|
@%p10 bra BB0_2;
|
|
|
|
ld.u8 %rs5, [%rd4+1];
|
|
and.b16 %rs7, %rs3, 255;
|
|
cvt.rn.f32.u16 %f340, %rs7;
|
|
div.rn.f32 %f341, %f340, 0f437F0000;
|
|
fma.rn.f32 %f342, %f341, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f343, %rs5;
|
|
div.rn.f32 %f344, %f343, 0f437F0000;
|
|
fma.rn.f32 %f345, %f344, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f346, %rs2;
|
|
div.rn.f32 %f347, %f346, 0f437F0000;
|
|
fma.rn.f32 %f348, %f347, 0f40000000, 0fBF800000;
|
|
mul.f32 %f349, %f345, %f345;
|
|
fma.rn.f32 %f350, %f342, %f342, %f349;
|
|
fma.rn.f32 %f351, %f348, %f348, %f350;
|
|
sqrt.rn.f32 %f352, %f351;
|
|
rcp.rn.f32 %f353, %f352;
|
|
mul.f32 %f354, %f342, %f353;
|
|
mul.f32 %f355, %f345, %f353;
|
|
mul.f32 %f356, %f348, %f353;
|
|
setp.eq.f32 %p70, %f354, 0f00000000;
|
|
setp.eq.f32 %p71, %f355, 0f00000000;
|
|
setp.eq.f32 %p72, %f356, 0f00000000;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r13, %r14}, [pixelID];
|
|
ld.global.v2.u32 {%r16, %r17}, [tileInfo];
|
|
add.s32 %r2, %r13, %r16;
|
|
add.s32 %r3, %r14, %r17;
|
|
and.pred %p11, %p70, %p71;
|
|
and.pred %p12, %p11, %p72;
|
|
@%p12 bra BB0_93;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_93:
|
|
cvt.u64.u32 %rd74, %r2;
|
|
cvt.u64.u32 %rd75, %r3;
|
|
mov.u64 %rd78, image_HDR;
|
|
cvta.global.u64 %rd73, %rd78;
|
|
mov.u32 %r127, 16;
|
|
// inline asm
|
|
call (%rd72), _rt_buffer_get_64, (%rd73, %r6, %r127, %rd74, %rd75, %rd9, %rd9);
|
|
// inline asm
|
|
mov.f32 %f1096, 0f00000000;
|
|
st.v4.f32 [%rd72], {%f1096, %f1096, %f1096, %f1096};
|
|
bra.uni BB0_94;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r30, %r31}, [pixelID];
|
|
cvt.u64.u32 %rd13, %r30;
|
|
cvt.u64.u32 %rd14, %r31;
|
|
mov.u64 %rd30, uvpos;
|
|
cvta.global.u64 %rd12, %rd30;
|
|
mov.u32 %r29, 12;
|
|
// inline asm
|
|
call (%rd11), _rt_buffer_get_64, (%rd12, %r6, %r29, %rd13, %rd14, %rd9, %rd9);
|
|
// inline asm
|
|
ld.f32 %f1, [%rd11];
|
|
ld.global.v2.u32 {%r34, %r35}, [pixelID];
|
|
cvt.u64.u32 %rd19, %r34;
|
|
cvt.u64.u32 %rd20, %r35;
|
|
// inline asm
|
|
call (%rd17), _rt_buffer_get_64, (%rd12, %r6, %r29, %rd19, %rd20, %rd9, %rd9);
|
|
// inline asm
|
|
ld.f32 %f2, [%rd17+4];
|
|
ld.global.v2.u32 {%r38, %r39}, [pixelID];
|
|
cvt.u64.u32 %rd25, %r38;
|
|
cvt.u64.u32 %rd26, %r39;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd12, %r6, %r29, %rd25, %rd26, %rd9, %rd9);
|
|
// inline asm
|
|
mul.f32 %f366, %f1, 0f3456BF95;
|
|
mul.f32 %f367, %f2, 0f3456BF95;
|
|
ld.f32 %f3, [%rd23+8];
|
|
mul.f32 %f368, %f3, 0f3456BF95;
|
|
abs.f32 %f369, %f366;
|
|
abs.f32 %f370, %f367;
|
|
abs.f32 %f371, %f368;
|
|
max.f32 %f372, %f369, %f370;
|
|
max.f32 %f373, %f372, %f371;
|
|
mov.f32 %f374, 0f38D1B717;
|
|
max.f32 %f4, %f373, %f374;
|
|
add.f32 %f5, %f4, %f4;
|
|
ld.global.f32 %f375, [voxelSize];
|
|
ld.global.f32 %f376, [voxelSize+4];
|
|
mul.f32 %f377, %f376, 0f00000000;
|
|
sub.f32 %f378, %f377, %f375;
|
|
ld.global.f32 %f379, [voxelSize+8];
|
|
fma.rn.f32 %f380, %f379, 0f00000000, %f378;
|
|
abs.f32 %f364, %f380;
|
|
add.u64 %rd29, %SP, 0;
|
|
cvta.to.local.u64 %rd31, %rd29;
|
|
mov.u32 %r28, 0;
|
|
st.local.u32 [%rd31+8], %r28;
|
|
st.local.u32 [%rd31+4], %r28;
|
|
st.local.u32 [%rd31], %r28;
|
|
ld.global.u32 %r27, [root];
|
|
mov.f32 %f360, 0fBF800000;
|
|
mov.f32 %f363, 0f00000000;
|
|
// inline asm
|
|
call _rt_trace_64, (%r27, %f1, %f2, %f3, %f360, %f363, %f363, %r28, %f363, %f364, %rd29, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f6, [%rd31];
|
|
abs.f32 %f381, %f6;
|
|
ld.local.f32 %f7, [%rd31+4];
|
|
abs.f32 %f382, %f7;
|
|
add.f32 %f383, %f381, %f382;
|
|
ld.local.f32 %f8, [%rd31+8];
|
|
abs.f32 %f384, %f8;
|
|
add.f32 %f385, %f383, %f384;
|
|
mov.f32 %f1141, 0f47C34F80;
|
|
setp.eq.f32 %p13, %f385, 0f00000000;
|
|
@%p13 bra BB0_4;
|
|
|
|
mul.f32 %f387, %f7, %f7;
|
|
fma.rn.f32 %f388, %f6, %f6, %f387;
|
|
fma.rn.f32 %f389, %f8, %f8, %f388;
|
|
sqrt.rn.f32 %f390, %f389;
|
|
max.f32 %f9, %f390, %f4;
|
|
rcp.rn.f32 %f391, %f390;
|
|
mul.f32 %f392, %f391, %f6;
|
|
mul.f32 %f393, %f391, %f7;
|
|
mul.f32 %f394, %f391, %f8;
|
|
fma.rn.f32 %f395, %f4, %f392, %f1;
|
|
fma.rn.f32 %f396, %f4, %f393, %f2;
|
|
fma.rn.f32 %f397, %f4, %f394, %f3;
|
|
add.f32 %f398, %f5, %f9;
|
|
sub.f32 %f10, %f395, %f398;
|
|
fma.rn.f32 %f11, %f398, 0f00000000, %f396;
|
|
fma.rn.f32 %f12, %f398, 0f00000000, %f397;
|
|
setp.geu.f32 %p14, %f9, 0f47C34F80;
|
|
mov.f32 %f1138, %f1;
|
|
mov.f32 %f1139, %f2;
|
|
mov.f32 %f1140, %f3;
|
|
@%p14 bra BB0_7;
|
|
|
|
mov.f32 %f1138, %f10;
|
|
mov.f32 %f1139, %f11;
|
|
mov.f32 %f1140, %f12;
|
|
mov.f32 %f1141, %f9;
|
|
bra.uni BB0_7;
|
|
|
|
BB0_4:
|
|
mov.f32 %f1138, %f1;
|
|
mov.f32 %f1139, %f2;
|
|
mov.f32 %f1140, %f3;
|
|
|
|
BB0_7:
|
|
mov.u32 %r130, 0;
|
|
add.u64 %rd79, %SP, 0;
|
|
ld.global.f32 %f407, [voxelSize+4];
|
|
ld.global.f32 %f408, [voxelSize];
|
|
fma.rn.f32 %f409, %f407, 0f00000000, %f408;
|
|
ld.global.f32 %f410, [voxelSize+8];
|
|
fma.rn.f32 %f411, %f410, 0f00000000, %f409;
|
|
abs.f32 %f406, %f411;
|
|
st.local.u32 [%rd31+8], %r130;
|
|
st.local.u32 [%rd31+4], %r130;
|
|
st.local.u32 [%rd31], %r130;
|
|
ld.global.u32 %r42, [root];
|
|
mov.f32 %f402, 0f3F800000;
|
|
// inline asm
|
|
call _rt_trace_64, (%r42, %f1, %f2, %f3, %f402, %f363, %f363, %r130, %f363, %f406, %rd79, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f17, [%rd31];
|
|
abs.f32 %f412, %f17;
|
|
ld.local.f32 %f18, [%rd31+4];
|
|
abs.f32 %f413, %f18;
|
|
add.f32 %f414, %f412, %f413;
|
|
ld.local.f32 %f19, [%rd31+8];
|
|
abs.f32 %f415, %f19;
|
|
add.f32 %f416, %f414, %f415;
|
|
setp.eq.f32 %p15, %f416, 0f00000000;
|
|
@%p15 bra BB0_10;
|
|
|
|
mul.f32 %f417, %f18, %f18;
|
|
fma.rn.f32 %f418, %f17, %f17, %f417;
|
|
fma.rn.f32 %f419, %f19, %f19, %f418;
|
|
sqrt.rn.f32 %f420, %f419;
|
|
max.f32 %f20, %f420, %f4;
|
|
rcp.rn.f32 %f421, %f420;
|
|
mul.f32 %f422, %f421, %f17;
|
|
mul.f32 %f423, %f421, %f18;
|
|
mul.f32 %f424, %f421, %f19;
|
|
fma.rn.f32 %f425, %f4, %f422, %f1;
|
|
fma.rn.f32 %f426, %f4, %f423, %f2;
|
|
fma.rn.f32 %f427, %f4, %f424, %f3;
|
|
add.f32 %f428, %f5, %f20;
|
|
add.f32 %f21, %f425, %f428;
|
|
fma.rn.f32 %f22, %f428, 0f00000000, %f426;
|
|
fma.rn.f32 %f23, %f428, 0f00000000, %f427;
|
|
setp.geu.f32 %p16, %f20, %f1141;
|
|
@%p16 bra BB0_10;
|
|
|
|
mov.f32 %f1138, %f21;
|
|
mov.f32 %f1139, %f22;
|
|
mov.f32 %f1140, %f23;
|
|
mov.f32 %f1141, %f20;
|
|
|
|
BB0_10:
|
|
mov.u32 %r131, 0;
|
|
add.u64 %rd86, %SP, 0;
|
|
cvta.to.local.u64 %rd85, %rd86;
|
|
add.u64 %rd80, %SP, 0;
|
|
ld.global.f32 %f437, [voxelSize];
|
|
ld.global.f32 %f438, [voxelSize+4];
|
|
neg.f32 %f439, %f438;
|
|
fma.rn.f32 %f440, %f437, 0f00000000, %f439;
|
|
ld.global.f32 %f441, [voxelSize+8];
|
|
fma.rn.f32 %f442, %f441, 0f00000000, %f440;
|
|
abs.f32 %f436, %f442;
|
|
st.local.u32 [%rd85+8], %r131;
|
|
st.local.u32 [%rd85+4], %r131;
|
|
st.local.u32 [%rd85], %r131;
|
|
ld.global.u32 %r45, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r45, %f1, %f2, %f3, %f363, %f360, %f363, %r131, %f363, %f436, %rd80, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f28, [%rd85];
|
|
abs.f32 %f443, %f28;
|
|
ld.local.f32 %f29, [%rd85+4];
|
|
abs.f32 %f444, %f29;
|
|
add.f32 %f445, %f443, %f444;
|
|
ld.local.f32 %f30, [%rd85+8];
|
|
abs.f32 %f446, %f30;
|
|
add.f32 %f447, %f445, %f446;
|
|
setp.eq.f32 %p17, %f447, 0f00000000;
|
|
@%p17 bra BB0_13;
|
|
|
|
mul.f32 %f448, %f29, %f29;
|
|
fma.rn.f32 %f449, %f28, %f28, %f448;
|
|
fma.rn.f32 %f450, %f30, %f30, %f449;
|
|
sqrt.rn.f32 %f451, %f450;
|
|
max.f32 %f31, %f451, %f4;
|
|
rcp.rn.f32 %f452, %f451;
|
|
mul.f32 %f453, %f452, %f28;
|
|
mul.f32 %f454, %f452, %f29;
|
|
mul.f32 %f455, %f452, %f30;
|
|
fma.rn.f32 %f456, %f4, %f453, %f1;
|
|
fma.rn.f32 %f457, %f4, %f454, %f2;
|
|
fma.rn.f32 %f458, %f4, %f455, %f3;
|
|
add.f32 %f459, %f5, %f31;
|
|
fma.rn.f32 %f32, %f459, 0f00000000, %f456;
|
|
sub.f32 %f33, %f457, %f459;
|
|
fma.rn.f32 %f34, %f459, 0f00000000, %f458;
|
|
setp.geu.f32 %p18, %f31, %f1141;
|
|
@%p18 bra BB0_13;
|
|
|
|
mov.f32 %f1138, %f32;
|
|
mov.f32 %f1139, %f33;
|
|
mov.f32 %f1140, %f34;
|
|
mov.f32 %f1141, %f31;
|
|
|
|
BB0_13:
|
|
mov.u32 %r132, 0;
|
|
add.u64 %rd88, %SP, 0;
|
|
cvta.to.local.u64 %rd87, %rd88;
|
|
add.u64 %rd81, %SP, 0;
|
|
ld.global.f32 %f468, [voxelSize];
|
|
ld.global.f32 %f469, [voxelSize+4];
|
|
fma.rn.f32 %f470, %f468, 0f00000000, %f469;
|
|
ld.global.f32 %f471, [voxelSize+8];
|
|
fma.rn.f32 %f472, %f471, 0f00000000, %f470;
|
|
abs.f32 %f467, %f472;
|
|
st.local.u32 [%rd87+8], %r132;
|
|
st.local.u32 [%rd87+4], %r132;
|
|
st.local.u32 [%rd87], %r132;
|
|
ld.global.u32 %r48, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r48, %f1, %f2, %f3, %f363, %f402, %f363, %r132, %f363, %f467, %rd81, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f39, [%rd87];
|
|
abs.f32 %f473, %f39;
|
|
ld.local.f32 %f40, [%rd87+4];
|
|
abs.f32 %f474, %f40;
|
|
add.f32 %f475, %f473, %f474;
|
|
ld.local.f32 %f41, [%rd87+8];
|
|
abs.f32 %f476, %f41;
|
|
add.f32 %f477, %f475, %f476;
|
|
setp.eq.f32 %p19, %f477, 0f00000000;
|
|
@%p19 bra BB0_16;
|
|
|
|
mul.f32 %f478, %f40, %f40;
|
|
fma.rn.f32 %f479, %f39, %f39, %f478;
|
|
fma.rn.f32 %f480, %f41, %f41, %f479;
|
|
sqrt.rn.f32 %f481, %f480;
|
|
max.f32 %f42, %f481, %f4;
|
|
rcp.rn.f32 %f482, %f481;
|
|
mul.f32 %f483, %f482, %f39;
|
|
mul.f32 %f484, %f482, %f40;
|
|
mul.f32 %f485, %f482, %f41;
|
|
fma.rn.f32 %f486, %f4, %f483, %f1;
|
|
fma.rn.f32 %f487, %f4, %f484, %f2;
|
|
fma.rn.f32 %f488, %f4, %f485, %f3;
|
|
add.f32 %f489, %f5, %f42;
|
|
fma.rn.f32 %f43, %f489, 0f00000000, %f486;
|
|
add.f32 %f44, %f487, %f489;
|
|
fma.rn.f32 %f45, %f489, 0f00000000, %f488;
|
|
setp.geu.f32 %p20, %f42, %f1141;
|
|
@%p20 bra BB0_16;
|
|
|
|
mov.f32 %f1138, %f43;
|
|
mov.f32 %f1139, %f44;
|
|
mov.f32 %f1140, %f45;
|
|
mov.f32 %f1141, %f42;
|
|
|
|
BB0_16:
|
|
mov.u32 %r133, 0;
|
|
add.u64 %rd90, %SP, 0;
|
|
cvta.to.local.u64 %rd89, %rd90;
|
|
add.u64 %rd82, %SP, 0;
|
|
ld.global.f32 %f498, [voxelSize];
|
|
ld.global.f32 %f499, [voxelSize+4];
|
|
mul.f32 %f500, %f499, 0f00000000;
|
|
fma.rn.f32 %f501, %f498, 0f00000000, %f500;
|
|
ld.global.f32 %f502, [voxelSize+8];
|
|
sub.f32 %f503, %f501, %f502;
|
|
abs.f32 %f497, %f503;
|
|
st.local.u32 [%rd89+8], %r133;
|
|
st.local.u32 [%rd89+4], %r133;
|
|
st.local.u32 [%rd89], %r133;
|
|
ld.global.u32 %r51, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r51, %f1, %f2, %f3, %f363, %f363, %f360, %r133, %f363, %f497, %rd82, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f50, [%rd89];
|
|
abs.f32 %f504, %f50;
|
|
ld.local.f32 %f51, [%rd89+4];
|
|
abs.f32 %f505, %f51;
|
|
add.f32 %f506, %f504, %f505;
|
|
ld.local.f32 %f52, [%rd89+8];
|
|
abs.f32 %f507, %f52;
|
|
add.f32 %f508, %f506, %f507;
|
|
setp.eq.f32 %p21, %f508, 0f00000000;
|
|
@%p21 bra BB0_19;
|
|
|
|
mul.f32 %f509, %f51, %f51;
|
|
fma.rn.f32 %f510, %f50, %f50, %f509;
|
|
fma.rn.f32 %f511, %f52, %f52, %f510;
|
|
sqrt.rn.f32 %f512, %f511;
|
|
max.f32 %f53, %f512, %f4;
|
|
rcp.rn.f32 %f513, %f512;
|
|
mul.f32 %f514, %f513, %f50;
|
|
mul.f32 %f515, %f513, %f51;
|
|
mul.f32 %f516, %f513, %f52;
|
|
fma.rn.f32 %f517, %f4, %f514, %f1;
|
|
fma.rn.f32 %f518, %f4, %f515, %f2;
|
|
fma.rn.f32 %f519, %f4, %f516, %f3;
|
|
add.f32 %f520, %f5, %f53;
|
|
fma.rn.f32 %f54, %f520, 0f00000000, %f517;
|
|
fma.rn.f32 %f55, %f520, 0f00000000, %f518;
|
|
sub.f32 %f56, %f519, %f520;
|
|
setp.geu.f32 %p22, %f53, %f1141;
|
|
@%p22 bra BB0_19;
|
|
|
|
mov.f32 %f1138, %f54;
|
|
mov.f32 %f1139, %f55;
|
|
mov.f32 %f1140, %f56;
|
|
mov.f32 %f1141, %f53;
|
|
|
|
BB0_19:
|
|
mov.u32 %r134, 0;
|
|
add.u64 %rd92, %SP, 0;
|
|
cvta.to.local.u64 %rd91, %rd92;
|
|
add.u64 %rd83, %SP, 0;
|
|
ld.global.f32 %f529, [voxelSize];
|
|
ld.global.f32 %f530, [voxelSize+4];
|
|
mul.f32 %f531, %f530, 0f00000000;
|
|
fma.rn.f32 %f532, %f529, 0f00000000, %f531;
|
|
ld.global.f32 %f533, [voxelSize+8];
|
|
add.f32 %f534, %f532, %f533;
|
|
abs.f32 %f528, %f534;
|
|
st.local.u32 [%rd91+8], %r134;
|
|
st.local.u32 [%rd91+4], %r134;
|
|
st.local.u32 [%rd91], %r134;
|
|
ld.global.u32 %r54, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r54, %f1, %f2, %f3, %f363, %f363, %f402, %r134, %f363, %f528, %rd83, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f61, [%rd91];
|
|
abs.f32 %f535, %f61;
|
|
ld.local.f32 %f62, [%rd91+4];
|
|
abs.f32 %f536, %f62;
|
|
add.f32 %f537, %f535, %f536;
|
|
ld.local.f32 %f63, [%rd91+8];
|
|
abs.f32 %f538, %f63;
|
|
add.f32 %f539, %f537, %f538;
|
|
setp.eq.f32 %p23, %f539, 0f00000000;
|
|
@%p23 bra BB0_22;
|
|
|
|
mul.f32 %f540, %f62, %f62;
|
|
fma.rn.f32 %f541, %f61, %f61, %f540;
|
|
fma.rn.f32 %f542, %f63, %f63, %f541;
|
|
sqrt.rn.f32 %f543, %f542;
|
|
max.f32 %f544, %f543, %f4;
|
|
rcp.rn.f32 %f545, %f543;
|
|
mul.f32 %f546, %f545, %f61;
|
|
mul.f32 %f547, %f545, %f62;
|
|
mul.f32 %f548, %f545, %f63;
|
|
fma.rn.f32 %f549, %f4, %f546, %f1;
|
|
fma.rn.f32 %f550, %f4, %f547, %f2;
|
|
fma.rn.f32 %f551, %f4, %f548, %f3;
|
|
add.f32 %f552, %f5, %f544;
|
|
fma.rn.f32 %f64, %f552, 0f00000000, %f549;
|
|
fma.rn.f32 %f65, %f552, 0f00000000, %f550;
|
|
add.f32 %f66, %f551, %f552;
|
|
setp.geu.f32 %p24, %f544, %f1141;
|
|
@%p24 bra BB0_22;
|
|
|
|
mov.f32 %f1138, %f64;
|
|
mov.f32 %f1139, %f65;
|
|
mov.f32 %f1140, %f66;
|
|
|
|
BB0_22:
|
|
mov.u32 %r136, 0;
|
|
neg.f32 %f70, %f4;
|
|
mul.f32 %f73, %f4, 0f00000000;
|
|
mul.f32 %f72, %f4, 0fBF3504F3;
|
|
mul.f32 %f74, %f4, 0f3F3504F3;
|
|
mul.f32 %f75, %f4, 0fBF13CD3A;
|
|
mul.f32 %f76, %f4, 0f3F13CD3A;
|
|
add.u64 %rd42, %SP, 12;
|
|
cvta.to.local.u64 %rd2, %rd42;
|
|
|
|
BB0_23:
|
|
mov.u32 %r129, 0;
|
|
ld.global.f32 %f562, [voxelSize];
|
|
ld.global.f32 %f563, [voxelSize+4];
|
|
mul.f32 %f564, %f563, 0f00000000;
|
|
sub.f32 %f565, %f564, %f562;
|
|
ld.global.f32 %f566, [voxelSize+8];
|
|
fma.rn.f32 %f567, %f566, 0f00000000, %f565;
|
|
abs.f32 %f560, %f567;
|
|
st.local.u32 [%rd2+8], %r129;
|
|
st.local.u32 [%rd2+4], %r129;
|
|
st.local.u32 [%rd2], %r129;
|
|
ld.global.u32 %r58, [root];
|
|
sub.f32 %f555, %f1140, %f73;
|
|
sub.f32 %f554, %f1139, %f73;
|
|
sub.f32 %f553, %f1138, %f70;
|
|
// inline asm
|
|
call _rt_trace_64, (%r58, %f553, %f554, %f555, %f360, %f363, %f363, %r129, %f363, %f560, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f82, [%rd2];
|
|
abs.f32 %f568, %f82;
|
|
ld.local.f32 %f83, [%rd2+4];
|
|
abs.f32 %f569, %f83;
|
|
add.f32 %f570, %f568, %f569;
|
|
ld.local.f32 %f84, [%rd2+8];
|
|
abs.f32 %f571, %f84;
|
|
add.f32 %f572, %f570, %f571;
|
|
mov.f32 %f1167, 0f47C34F80;
|
|
setp.eq.f32 %p25, %f572, 0f00000000;
|
|
@%p25 bra BB0_24;
|
|
|
|
mul.f32 %f574, %f83, %f83;
|
|
fma.rn.f32 %f575, %f82, %f82, %f574;
|
|
fma.rn.f32 %f576, %f84, %f84, %f575;
|
|
sqrt.rn.f32 %f577, %f576;
|
|
max.f32 %f85, %f577, %f4;
|
|
add.f32 %f578, %f5, %f85;
|
|
sub.f32 %f86, %f1138, %f578;
|
|
fma.rn.f32 %f87, %f578, 0f00000000, %f1139;
|
|
fma.rn.f32 %f88, %f578, 0f00000000, %f1140;
|
|
setp.geu.f32 %p26, %f85, 0f47C34F80;
|
|
mov.f32 %f1164, %f1138;
|
|
mov.f32 %f1165, %f1139;
|
|
mov.f32 %f1166, %f1140;
|
|
@%p26 bra BB0_27;
|
|
|
|
mov.f32 %f1164, %f86;
|
|
mov.f32 %f1165, %f87;
|
|
mov.f32 %f1166, %f88;
|
|
mov.f32 %f1167, %f85;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_24:
|
|
mov.f32 %f1164, %f1138;
|
|
mov.f32 %f1165, %f1139;
|
|
mov.f32 %f1166, %f1140;
|
|
|
|
BB0_27:
|
|
ld.global.f32 %f587, [voxelSize+4];
|
|
ld.global.f32 %f588, [voxelSize];
|
|
fma.rn.f32 %f589, %f587, 0f00000000, %f588;
|
|
ld.global.f32 %f590, [voxelSize+8];
|
|
fma.rn.f32 %f591, %f590, 0f00000000, %f589;
|
|
abs.f32 %f586, %f591;
|
|
mov.u32 %r62, 0;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r61, [root];
|
|
sub.f32 %f579, %f1138, %f4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r61, %f579, %f554, %f555, %f402, %f363, %f363, %r62, %f363, %f586, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f93, [%rd2];
|
|
abs.f32 %f592, %f93;
|
|
ld.local.f32 %f94, [%rd2+4];
|
|
abs.f32 %f593, %f94;
|
|
add.f32 %f594, %f592, %f593;
|
|
ld.local.f32 %f95, [%rd2+8];
|
|
abs.f32 %f595, %f95;
|
|
add.f32 %f596, %f594, %f595;
|
|
setp.eq.f32 %p27, %f596, 0f00000000;
|
|
@%p27 bra BB0_30;
|
|
|
|
mul.f32 %f597, %f94, %f94;
|
|
fma.rn.f32 %f598, %f93, %f93, %f597;
|
|
fma.rn.f32 %f599, %f95, %f95, %f598;
|
|
sqrt.rn.f32 %f600, %f599;
|
|
max.f32 %f96, %f600, %f4;
|
|
add.f32 %f97, %f5, %f96;
|
|
fma.rn.f32 %f98, %f97, 0f00000000, %f1139;
|
|
fma.rn.f32 %f99, %f97, 0f00000000, %f1140;
|
|
setp.geu.f32 %p28, %f96, %f1167;
|
|
@%p28 bra BB0_30;
|
|
|
|
add.f32 %f1164, %f1138, %f97;
|
|
mov.f32 %f1165, %f98;
|
|
mov.f32 %f1166, %f99;
|
|
mov.f32 %f1167, %f96;
|
|
|
|
BB0_30:
|
|
ld.global.f32 %f609, [voxelSize];
|
|
ld.global.f32 %f610, [voxelSize+4];
|
|
neg.f32 %f611, %f610;
|
|
fma.rn.f32 %f612, %f609, 0f00000000, %f611;
|
|
ld.global.f32 %f613, [voxelSize+8];
|
|
fma.rn.f32 %f614, %f613, 0f00000000, %f612;
|
|
abs.f32 %f608, %f614;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r64, [root];
|
|
sub.f32 %f602, %f1139, %f70;
|
|
sub.f32 %f601, %f1138, %f73;
|
|
// inline asm
|
|
call _rt_trace_64, (%r64, %f601, %f602, %f555, %f363, %f360, %f363, %r62, %f363, %f608, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f106, [%rd2];
|
|
abs.f32 %f615, %f106;
|
|
ld.local.f32 %f107, [%rd2+4];
|
|
abs.f32 %f616, %f107;
|
|
add.f32 %f617, %f615, %f616;
|
|
ld.local.f32 %f108, [%rd2+8];
|
|
abs.f32 %f618, %f108;
|
|
add.f32 %f619, %f617, %f618;
|
|
setp.eq.f32 %p29, %f619, 0f00000000;
|
|
@%p29 bra BB0_33;
|
|
|
|
mul.f32 %f620, %f107, %f107;
|
|
fma.rn.f32 %f621, %f106, %f106, %f620;
|
|
fma.rn.f32 %f622, %f108, %f108, %f621;
|
|
sqrt.rn.f32 %f623, %f622;
|
|
max.f32 %f109, %f623, %f4;
|
|
add.f32 %f624, %f5, %f109;
|
|
fma.rn.f32 %f110, %f624, 0f00000000, %f1138;
|
|
sub.f32 %f111, %f1139, %f624;
|
|
fma.rn.f32 %f112, %f624, 0f00000000, %f1140;
|
|
setp.geu.f32 %p30, %f109, %f1167;
|
|
@%p30 bra BB0_33;
|
|
|
|
mov.f32 %f1164, %f110;
|
|
mov.f32 %f1165, %f111;
|
|
mov.f32 %f1166, %f112;
|
|
mov.f32 %f1167, %f109;
|
|
|
|
BB0_33:
|
|
ld.global.f32 %f633, [voxelSize];
|
|
ld.global.f32 %f634, [voxelSize+4];
|
|
fma.rn.f32 %f635, %f633, 0f00000000, %f634;
|
|
ld.global.f32 %f636, [voxelSize+8];
|
|
fma.rn.f32 %f637, %f636, 0f00000000, %f635;
|
|
abs.f32 %f632, %f637;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r67, [root];
|
|
sub.f32 %f626, %f1139, %f4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r67, %f601, %f626, %f555, %f363, %f402, %f363, %r62, %f363, %f632, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f117, [%rd2];
|
|
abs.f32 %f638, %f117;
|
|
ld.local.f32 %f118, [%rd2+4];
|
|
abs.f32 %f639, %f118;
|
|
add.f32 %f640, %f638, %f639;
|
|
ld.local.f32 %f119, [%rd2+8];
|
|
abs.f32 %f641, %f119;
|
|
add.f32 %f642, %f640, %f641;
|
|
setp.eq.f32 %p31, %f642, 0f00000000;
|
|
@%p31 bra BB0_36;
|
|
|
|
mul.f32 %f643, %f118, %f118;
|
|
fma.rn.f32 %f644, %f117, %f117, %f643;
|
|
fma.rn.f32 %f645, %f119, %f119, %f644;
|
|
sqrt.rn.f32 %f646, %f645;
|
|
max.f32 %f120, %f646, %f4;
|
|
add.f32 %f121, %f5, %f120;
|
|
fma.rn.f32 %f122, %f121, 0f00000000, %f1138;
|
|
fma.rn.f32 %f123, %f121, 0f00000000, %f1140;
|
|
setp.geu.f32 %p32, %f120, %f1167;
|
|
@%p32 bra BB0_36;
|
|
|
|
add.f32 %f1165, %f1139, %f121;
|
|
mov.f32 %f1164, %f122;
|
|
mov.f32 %f1166, %f123;
|
|
mov.f32 %f1167, %f120;
|
|
|
|
BB0_36:
|
|
ld.global.f32 %f655, [voxelSize];
|
|
ld.global.f32 %f656, [voxelSize+4];
|
|
mul.f32 %f657, %f656, 0f00000000;
|
|
fma.rn.f32 %f658, %f655, 0f00000000, %f657;
|
|
ld.global.f32 %f659, [voxelSize+8];
|
|
sub.f32 %f660, %f658, %f659;
|
|
abs.f32 %f654, %f660;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r70, [root];
|
|
sub.f32 %f649, %f1140, %f70;
|
|
// inline asm
|
|
call _rt_trace_64, (%r70, %f601, %f554, %f649, %f363, %f363, %f360, %r62, %f363, %f654, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f129, [%rd2];
|
|
abs.f32 %f661, %f129;
|
|
ld.local.f32 %f130, [%rd2+4];
|
|
abs.f32 %f662, %f130;
|
|
add.f32 %f663, %f661, %f662;
|
|
ld.local.f32 %f131, [%rd2+8];
|
|
abs.f32 %f664, %f131;
|
|
add.f32 %f665, %f663, %f664;
|
|
setp.eq.f32 %p33, %f665, 0f00000000;
|
|
@%p33 bra BB0_39;
|
|
|
|
mul.f32 %f666, %f130, %f130;
|
|
fma.rn.f32 %f667, %f129, %f129, %f666;
|
|
fma.rn.f32 %f668, %f131, %f131, %f667;
|
|
sqrt.rn.f32 %f669, %f668;
|
|
max.f32 %f132, %f669, %f4;
|
|
add.f32 %f670, %f5, %f132;
|
|
fma.rn.f32 %f133, %f670, 0f00000000, %f1138;
|
|
fma.rn.f32 %f134, %f670, 0f00000000, %f1139;
|
|
sub.f32 %f135, %f1140, %f670;
|
|
setp.geu.f32 %p34, %f132, %f1167;
|
|
@%p34 bra BB0_39;
|
|
|
|
mov.f32 %f1164, %f133;
|
|
mov.f32 %f1165, %f134;
|
|
mov.f32 %f1166, %f135;
|
|
mov.f32 %f1167, %f132;
|
|
|
|
BB0_39:
|
|
ld.global.f32 %f679, [voxelSize];
|
|
ld.global.f32 %f680, [voxelSize+4];
|
|
mul.f32 %f681, %f680, 0f00000000;
|
|
fma.rn.f32 %f682, %f679, 0f00000000, %f681;
|
|
ld.global.f32 %f683, [voxelSize+8];
|
|
add.f32 %f684, %f682, %f683;
|
|
abs.f32 %f678, %f684;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r73, [root];
|
|
sub.f32 %f673, %f1140, %f4;
|
|
// inline asm
|
|
call _rt_trace_64, (%r73, %f601, %f554, %f673, %f363, %f363, %f402, %r62, %f363, %f678, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f140, [%rd2];
|
|
abs.f32 %f685, %f140;
|
|
ld.local.f32 %f141, [%rd2+4];
|
|
abs.f32 %f686, %f141;
|
|
add.f32 %f687, %f685, %f686;
|
|
ld.local.f32 %f142, [%rd2+8];
|
|
abs.f32 %f688, %f142;
|
|
add.f32 %f689, %f687, %f688;
|
|
setp.eq.f32 %p35, %f689, 0f00000000;
|
|
@%p35 bra BB0_42;
|
|
|
|
mul.f32 %f690, %f141, %f141;
|
|
fma.rn.f32 %f691, %f140, %f140, %f690;
|
|
fma.rn.f32 %f692, %f142, %f142, %f691;
|
|
sqrt.rn.f32 %f693, %f692;
|
|
max.f32 %f143, %f693, %f4;
|
|
add.f32 %f144, %f5, %f143;
|
|
fma.rn.f32 %f145, %f144, 0f00000000, %f1138;
|
|
fma.rn.f32 %f146, %f144, 0f00000000, %f1139;
|
|
setp.geu.f32 %p36, %f143, %f1167;
|
|
@%p36 bra BB0_42;
|
|
|
|
add.f32 %f1166, %f1140, %f144;
|
|
mov.f32 %f1164, %f145;
|
|
mov.f32 %f1165, %f146;
|
|
mov.f32 %f1167, %f143;
|
|
|
|
BB0_42:
|
|
ld.global.f32 %f702, [voxelSize];
|
|
ld.global.f32 %f703, [voxelSize+4];
|
|
mul.f32 %f704, %f703, 0fBF3504F3;
|
|
fma.rn.f32 %f705, %f702, 0fBF3504F3, %f704;
|
|
ld.global.f32 %f706, [voxelSize+8];
|
|
fma.rn.f32 %f707, %f706, 0f00000000, %f705;
|
|
abs.f32 %f708, %f707;
|
|
mul.f32 %f701, %f708, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r76, [root];
|
|
sub.f32 %f695, %f1139, %f72;
|
|
sub.f32 %f694, %f1138, %f72;
|
|
mov.f32 %f698, 0fBF3504F3;
|
|
// inline asm
|
|
call _rt_trace_64, (%r76, %f694, %f695, %f555, %f698, %f698, %f363, %r62, %f363, %f701, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f155, [%rd2];
|
|
abs.f32 %f709, %f155;
|
|
ld.local.f32 %f156, [%rd2+4];
|
|
abs.f32 %f710, %f156;
|
|
add.f32 %f711, %f709, %f710;
|
|
ld.local.f32 %f157, [%rd2+8];
|
|
abs.f32 %f712, %f157;
|
|
add.f32 %f713, %f711, %f712;
|
|
setp.eq.f32 %p37, %f713, 0f00000000;
|
|
@%p37 bra BB0_45;
|
|
|
|
mul.f32 %f714, %f156, %f156;
|
|
fma.rn.f32 %f715, %f155, %f155, %f714;
|
|
fma.rn.f32 %f716, %f157, %f157, %f715;
|
|
sqrt.rn.f32 %f717, %f716;
|
|
max.f32 %f158, %f717, %f4;
|
|
add.f32 %f718, %f5, %f158;
|
|
fma.rn.f32 %f159, %f718, 0fBF3504F3, %f1138;
|
|
fma.rn.f32 %f160, %f718, 0fBF3504F3, %f1139;
|
|
fma.rn.f32 %f161, %f718, 0f00000000, %f1140;
|
|
setp.geu.f32 %p38, %f158, %f1167;
|
|
@%p38 bra BB0_45;
|
|
|
|
mov.f32 %f1164, %f159;
|
|
mov.f32 %f1165, %f160;
|
|
mov.f32 %f1166, %f161;
|
|
mov.f32 %f1167, %f158;
|
|
|
|
BB0_45:
|
|
ld.global.f32 %f727, [voxelSize];
|
|
ld.global.f32 %f728, [voxelSize+4];
|
|
mul.f32 %f729, %f728, 0f3F3504F3;
|
|
fma.rn.f32 %f730, %f727, 0fBF3504F3, %f729;
|
|
ld.global.f32 %f731, [voxelSize+8];
|
|
fma.rn.f32 %f732, %f731, 0f00000000, %f730;
|
|
abs.f32 %f733, %f732;
|
|
mul.f32 %f726, %f733, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r79, [root];
|
|
sub.f32 %f720, %f1139, %f74;
|
|
mov.f32 %f723, 0f3F3504F3;
|
|
// inline asm
|
|
call _rt_trace_64, (%r79, %f694, %f720, %f555, %f698, %f723, %f363, %r62, %f363, %f726, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f167, [%rd2];
|
|
abs.f32 %f734, %f167;
|
|
ld.local.f32 %f168, [%rd2+4];
|
|
abs.f32 %f735, %f168;
|
|
add.f32 %f736, %f734, %f735;
|
|
ld.local.f32 %f169, [%rd2+8];
|
|
abs.f32 %f737, %f169;
|
|
add.f32 %f738, %f736, %f737;
|
|
setp.eq.f32 %p39, %f738, 0f00000000;
|
|
@%p39 bra BB0_48;
|
|
|
|
mul.f32 %f739, %f168, %f168;
|
|
fma.rn.f32 %f740, %f167, %f167, %f739;
|
|
fma.rn.f32 %f741, %f169, %f169, %f740;
|
|
sqrt.rn.f32 %f742, %f741;
|
|
max.f32 %f170, %f742, %f4;
|
|
add.f32 %f743, %f5, %f170;
|
|
fma.rn.f32 %f171, %f743, 0fBF3504F3, %f1138;
|
|
fma.rn.f32 %f172, %f743, 0f3F3504F3, %f1139;
|
|
fma.rn.f32 %f173, %f743, 0f00000000, %f1140;
|
|
setp.geu.f32 %p40, %f170, %f1167;
|
|
@%p40 bra BB0_48;
|
|
|
|
mov.f32 %f1164, %f171;
|
|
mov.f32 %f1165, %f172;
|
|
mov.f32 %f1166, %f173;
|
|
mov.f32 %f1167, %f170;
|
|
|
|
BB0_48:
|
|
sub.f32 %f1119, %f1139, %f74;
|
|
mov.f32 %f1118, 0f3F3504F3;
|
|
ld.global.f32 %f752, [voxelSize];
|
|
ld.global.f32 %f753, [voxelSize+4];
|
|
mul.f32 %f754, %f753, 0f3F3504F3;
|
|
fma.rn.f32 %f755, %f752, 0f3F3504F3, %f754;
|
|
ld.global.f32 %f756, [voxelSize+8];
|
|
fma.rn.f32 %f757, %f756, 0f00000000, %f755;
|
|
abs.f32 %f758, %f757;
|
|
mul.f32 %f751, %f758, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r82, [root];
|
|
sub.f32 %f744, %f1138, %f74;
|
|
// inline asm
|
|
call _rt_trace_64, (%r82, %f744, %f1119, %f555, %f1118, %f1118, %f363, %r62, %f363, %f751, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f179, [%rd2];
|
|
abs.f32 %f759, %f179;
|
|
ld.local.f32 %f180, [%rd2+4];
|
|
abs.f32 %f760, %f180;
|
|
add.f32 %f761, %f759, %f760;
|
|
ld.local.f32 %f181, [%rd2+8];
|
|
abs.f32 %f762, %f181;
|
|
add.f32 %f763, %f761, %f762;
|
|
setp.eq.f32 %p41, %f763, 0f00000000;
|
|
@%p41 bra BB0_51;
|
|
|
|
mul.f32 %f764, %f180, %f180;
|
|
fma.rn.f32 %f765, %f179, %f179, %f764;
|
|
fma.rn.f32 %f766, %f181, %f181, %f765;
|
|
sqrt.rn.f32 %f767, %f766;
|
|
max.f32 %f182, %f767, %f4;
|
|
add.f32 %f768, %f5, %f182;
|
|
fma.rn.f32 %f183, %f768, 0f3F3504F3, %f1138;
|
|
fma.rn.f32 %f184, %f768, 0f3F3504F3, %f1139;
|
|
fma.rn.f32 %f185, %f768, 0f00000000, %f1140;
|
|
setp.geu.f32 %p42, %f182, %f1167;
|
|
@%p42 bra BB0_51;
|
|
|
|
mov.f32 %f1164, %f183;
|
|
mov.f32 %f1165, %f184;
|
|
mov.f32 %f1166, %f185;
|
|
mov.f32 %f1167, %f182;
|
|
|
|
BB0_51:
|
|
sub.f32 %f1122, %f1139, %f72;
|
|
mov.f32 %f1121, 0f3F3504F3;
|
|
mov.f32 %f1120, 0fBF3504F3;
|
|
ld.global.f32 %f777, [voxelSize];
|
|
ld.global.f32 %f778, [voxelSize+4];
|
|
mul.f32 %f779, %f778, 0fBF3504F3;
|
|
fma.rn.f32 %f780, %f777, 0f3F3504F3, %f779;
|
|
ld.global.f32 %f781, [voxelSize+8];
|
|
fma.rn.f32 %f782, %f781, 0f00000000, %f780;
|
|
abs.f32 %f783, %f782;
|
|
mul.f32 %f776, %f783, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r85, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r85, %f744, %f1122, %f555, %f1121, %f1120, %f363, %r62, %f363, %f776, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f190, [%rd2];
|
|
abs.f32 %f784, %f190;
|
|
ld.local.f32 %f191, [%rd2+4];
|
|
abs.f32 %f785, %f191;
|
|
add.f32 %f786, %f784, %f785;
|
|
ld.local.f32 %f192, [%rd2+8];
|
|
abs.f32 %f787, %f192;
|
|
add.f32 %f788, %f786, %f787;
|
|
setp.eq.f32 %p43, %f788, 0f00000000;
|
|
@%p43 bra BB0_54;
|
|
|
|
mul.f32 %f789, %f191, %f191;
|
|
fma.rn.f32 %f790, %f190, %f190, %f789;
|
|
fma.rn.f32 %f791, %f192, %f192, %f790;
|
|
sqrt.rn.f32 %f792, %f791;
|
|
max.f32 %f193, %f792, %f4;
|
|
add.f32 %f793, %f5, %f193;
|
|
fma.rn.f32 %f194, %f793, 0f3F3504F3, %f1138;
|
|
fma.rn.f32 %f195, %f793, 0fBF3504F3, %f1139;
|
|
fma.rn.f32 %f196, %f793, 0f00000000, %f1140;
|
|
setp.geu.f32 %p44, %f193, %f1167;
|
|
@%p44 bra BB0_54;
|
|
|
|
mov.f32 %f1164, %f194;
|
|
mov.f32 %f1165, %f195;
|
|
mov.f32 %f1166, %f196;
|
|
mov.f32 %f1167, %f193;
|
|
|
|
BB0_54:
|
|
sub.f32 %f1124, %f1139, %f72;
|
|
mov.f32 %f1123, 0fBF3504F3;
|
|
ld.global.f32 %f802, [voxelSize];
|
|
ld.global.f32 %f803, [voxelSize+4];
|
|
mul.f32 %f804, %f803, 0fBF3504F3;
|
|
fma.rn.f32 %f805, %f802, 0f00000000, %f804;
|
|
ld.global.f32 %f806, [voxelSize+8];
|
|
fma.rn.f32 %f807, %f806, 0fBF3504F3, %f805;
|
|
abs.f32 %f808, %f807;
|
|
mul.f32 %f801, %f808, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r88, [root];
|
|
sub.f32 %f796, %f1140, %f72;
|
|
// inline asm
|
|
call _rt_trace_64, (%r88, %f601, %f1124, %f796, %f363, %f1123, %f1123, %r62, %f363, %f801, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f203, [%rd2];
|
|
abs.f32 %f809, %f203;
|
|
ld.local.f32 %f204, [%rd2+4];
|
|
abs.f32 %f810, %f204;
|
|
add.f32 %f811, %f809, %f810;
|
|
ld.local.f32 %f205, [%rd2+8];
|
|
abs.f32 %f812, %f205;
|
|
add.f32 %f813, %f811, %f812;
|
|
setp.eq.f32 %p45, %f813, 0f00000000;
|
|
@%p45 bra BB0_57;
|
|
|
|
mul.f32 %f814, %f204, %f204;
|
|
fma.rn.f32 %f815, %f203, %f203, %f814;
|
|
fma.rn.f32 %f816, %f205, %f205, %f815;
|
|
sqrt.rn.f32 %f817, %f816;
|
|
max.f32 %f206, %f817, %f4;
|
|
add.f32 %f818, %f5, %f206;
|
|
fma.rn.f32 %f207, %f818, 0f00000000, %f1138;
|
|
fma.rn.f32 %f208, %f818, 0fBF3504F3, %f1139;
|
|
fma.rn.f32 %f209, %f818, 0fBF3504F3, %f1140;
|
|
setp.geu.f32 %p46, %f206, %f1167;
|
|
@%p46 bra BB0_57;
|
|
|
|
mov.f32 %f1164, %f207;
|
|
mov.f32 %f1165, %f208;
|
|
mov.f32 %f1166, %f209;
|
|
mov.f32 %f1167, %f206;
|
|
|
|
BB0_57:
|
|
sub.f32 %f1127, %f1139, %f72;
|
|
mov.f32 %f1126, 0f3F3504F3;
|
|
mov.f32 %f1125, 0fBF3504F3;
|
|
ld.global.f32 %f827, [voxelSize];
|
|
ld.global.f32 %f828, [voxelSize+4];
|
|
mul.f32 %f829, %f828, 0fBF3504F3;
|
|
fma.rn.f32 %f830, %f827, 0f00000000, %f829;
|
|
ld.global.f32 %f831, [voxelSize+8];
|
|
fma.rn.f32 %f832, %f831, 0f3F3504F3, %f830;
|
|
abs.f32 %f833, %f832;
|
|
mul.f32 %f826, %f833, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r91, [root];
|
|
sub.f32 %f821, %f1140, %f74;
|
|
// inline asm
|
|
call _rt_trace_64, (%r91, %f601, %f1127, %f821, %f363, %f1125, %f1126, %r62, %f363, %f826, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f215, [%rd2];
|
|
abs.f32 %f834, %f215;
|
|
ld.local.f32 %f216, [%rd2+4];
|
|
abs.f32 %f835, %f216;
|
|
add.f32 %f836, %f834, %f835;
|
|
ld.local.f32 %f217, [%rd2+8];
|
|
abs.f32 %f837, %f217;
|
|
add.f32 %f838, %f836, %f837;
|
|
setp.eq.f32 %p47, %f838, 0f00000000;
|
|
@%p47 bra BB0_60;
|
|
|
|
mul.f32 %f839, %f216, %f216;
|
|
fma.rn.f32 %f840, %f215, %f215, %f839;
|
|
fma.rn.f32 %f841, %f217, %f217, %f840;
|
|
sqrt.rn.f32 %f842, %f841;
|
|
max.f32 %f218, %f842, %f4;
|
|
add.f32 %f843, %f5, %f218;
|
|
fma.rn.f32 %f219, %f843, 0f00000000, %f1138;
|
|
fma.rn.f32 %f220, %f843, 0fBF3504F3, %f1139;
|
|
fma.rn.f32 %f221, %f843, 0f3F3504F3, %f1140;
|
|
setp.geu.f32 %p48, %f218, %f1167;
|
|
@%p48 bra BB0_60;
|
|
|
|
mov.f32 %f1164, %f219;
|
|
mov.f32 %f1165, %f220;
|
|
mov.f32 %f1166, %f221;
|
|
mov.f32 %f1167, %f218;
|
|
|
|
BB0_60:
|
|
sub.f32 %f1129, %f1139, %f74;
|
|
mov.f32 %f1128, 0f3F3504F3;
|
|
ld.global.f32 %f852, [voxelSize];
|
|
ld.global.f32 %f853, [voxelSize+4];
|
|
mul.f32 %f854, %f853, 0f3F3504F3;
|
|
fma.rn.f32 %f855, %f852, 0f00000000, %f854;
|
|
ld.global.f32 %f856, [voxelSize+8];
|
|
fma.rn.f32 %f857, %f856, 0f3F3504F3, %f855;
|
|
abs.f32 %f858, %f857;
|
|
mul.f32 %f851, %f858, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r94, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r94, %f601, %f1129, %f821, %f363, %f1128, %f1128, %r62, %f363, %f851, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f226, [%rd2];
|
|
abs.f32 %f859, %f226;
|
|
ld.local.f32 %f227, [%rd2+4];
|
|
abs.f32 %f860, %f227;
|
|
add.f32 %f861, %f859, %f860;
|
|
ld.local.f32 %f228, [%rd2+8];
|
|
abs.f32 %f862, %f228;
|
|
add.f32 %f863, %f861, %f862;
|
|
setp.eq.f32 %p49, %f863, 0f00000000;
|
|
@%p49 bra BB0_63;
|
|
|
|
mul.f32 %f864, %f227, %f227;
|
|
fma.rn.f32 %f865, %f226, %f226, %f864;
|
|
fma.rn.f32 %f866, %f228, %f228, %f865;
|
|
sqrt.rn.f32 %f867, %f866;
|
|
max.f32 %f229, %f867, %f4;
|
|
add.f32 %f868, %f5, %f229;
|
|
fma.rn.f32 %f230, %f868, 0f00000000, %f1138;
|
|
fma.rn.f32 %f231, %f868, 0f3F3504F3, %f1139;
|
|
fma.rn.f32 %f232, %f868, 0f3F3504F3, %f1140;
|
|
setp.geu.f32 %p50, %f229, %f1167;
|
|
@%p50 bra BB0_63;
|
|
|
|
mov.f32 %f1164, %f230;
|
|
mov.f32 %f1165, %f231;
|
|
mov.f32 %f1166, %f232;
|
|
mov.f32 %f1167, %f229;
|
|
|
|
BB0_63:
|
|
sub.f32 %f1133, %f1140, %f72;
|
|
sub.f32 %f1132, %f1139, %f74;
|
|
mov.f32 %f1131, 0f3F3504F3;
|
|
mov.f32 %f1130, 0fBF3504F3;
|
|
ld.global.f32 %f877, [voxelSize];
|
|
ld.global.f32 %f878, [voxelSize+4];
|
|
mul.f32 %f879, %f878, 0f3F3504F3;
|
|
fma.rn.f32 %f880, %f877, 0f00000000, %f879;
|
|
ld.global.f32 %f881, [voxelSize+8];
|
|
fma.rn.f32 %f882, %f881, 0fBF3504F3, %f880;
|
|
abs.f32 %f883, %f882;
|
|
mul.f32 %f876, %f883, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r97, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r97, %f601, %f1132, %f1133, %f363, %f1131, %f1130, %r62, %f363, %f876, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f237, [%rd2];
|
|
abs.f32 %f884, %f237;
|
|
ld.local.f32 %f238, [%rd2+4];
|
|
abs.f32 %f885, %f238;
|
|
add.f32 %f886, %f884, %f885;
|
|
ld.local.f32 %f239, [%rd2+8];
|
|
abs.f32 %f887, %f239;
|
|
add.f32 %f888, %f886, %f887;
|
|
setp.eq.f32 %p51, %f888, 0f00000000;
|
|
@%p51 bra BB0_66;
|
|
|
|
mul.f32 %f889, %f238, %f238;
|
|
fma.rn.f32 %f890, %f237, %f237, %f889;
|
|
fma.rn.f32 %f891, %f239, %f239, %f890;
|
|
sqrt.rn.f32 %f892, %f891;
|
|
max.f32 %f240, %f892, %f4;
|
|
add.f32 %f893, %f5, %f240;
|
|
fma.rn.f32 %f241, %f893, 0f00000000, %f1138;
|
|
fma.rn.f32 %f242, %f893, 0f3F3504F3, %f1139;
|
|
fma.rn.f32 %f243, %f893, 0fBF3504F3, %f1140;
|
|
setp.geu.f32 %p52, %f240, %f1167;
|
|
@%p52 bra BB0_66;
|
|
|
|
mov.f32 %f1164, %f241;
|
|
mov.f32 %f1165, %f242;
|
|
mov.f32 %f1166, %f243;
|
|
mov.f32 %f1167, %f240;
|
|
|
|
BB0_66:
|
|
ld.global.f32 %f902, [voxelSize];
|
|
ld.global.f32 %f903, [voxelSize+4];
|
|
mul.f32 %f904, %f903, 0fBF13CD3A;
|
|
fma.rn.f32 %f905, %f902, 0fBF13CD3A, %f904;
|
|
ld.global.f32 %f906, [voxelSize+8];
|
|
fma.rn.f32 %f907, %f906, 0fBF13CD3A, %f905;
|
|
abs.f32 %f908, %f907;
|
|
mul.f32 %f901, %f908, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r100, [root];
|
|
sub.f32 %f896, %f1140, %f75;
|
|
sub.f32 %f895, %f1139, %f75;
|
|
sub.f32 %f894, %f1138, %f75;
|
|
mov.f32 %f899, 0fBF13CD3A;
|
|
// inline asm
|
|
call _rt_trace_64, (%r100, %f894, %f895, %f896, %f899, %f899, %f899, %r62, %f363, %f901, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f251, [%rd2];
|
|
abs.f32 %f909, %f251;
|
|
ld.local.f32 %f252, [%rd2+4];
|
|
abs.f32 %f910, %f252;
|
|
add.f32 %f911, %f909, %f910;
|
|
ld.local.f32 %f253, [%rd2+8];
|
|
abs.f32 %f912, %f253;
|
|
add.f32 %f913, %f911, %f912;
|
|
setp.eq.f32 %p53, %f913, 0f00000000;
|
|
@%p53 bra BB0_69;
|
|
|
|
mul.f32 %f914, %f252, %f252;
|
|
fma.rn.f32 %f915, %f251, %f251, %f914;
|
|
fma.rn.f32 %f916, %f253, %f253, %f915;
|
|
sqrt.rn.f32 %f917, %f916;
|
|
max.f32 %f254, %f917, %f4;
|
|
add.f32 %f918, %f5, %f254;
|
|
fma.rn.f32 %f255, %f918, 0fBF13CD3A, %f1138;
|
|
fma.rn.f32 %f256, %f918, 0fBF13CD3A, %f1139;
|
|
fma.rn.f32 %f257, %f918, 0fBF13CD3A, %f1140;
|
|
setp.geu.f32 %p54, %f254, %f1167;
|
|
@%p54 bra BB0_69;
|
|
|
|
mov.f32 %f1164, %f255;
|
|
mov.f32 %f1165, %f256;
|
|
mov.f32 %f1166, %f257;
|
|
mov.f32 %f1167, %f254;
|
|
|
|
BB0_69:
|
|
mov.f32 %f1111, 0fBF13CD3A;
|
|
ld.global.f32 %f927, [voxelSize];
|
|
ld.global.f32 %f928, [voxelSize+4];
|
|
mul.f32 %f929, %f928, 0fBF13CD3A;
|
|
fma.rn.f32 %f930, %f927, 0fBF13CD3A, %f929;
|
|
ld.global.f32 %f931, [voxelSize+8];
|
|
fma.rn.f32 %f932, %f931, 0f3F13CD3A, %f930;
|
|
abs.f32 %f933, %f932;
|
|
mul.f32 %f926, %f933, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r103, [root];
|
|
sub.f32 %f921, %f1140, %f76;
|
|
mov.f32 %f924, 0f3F13CD3A;
|
|
// inline asm
|
|
call _rt_trace_64, (%r103, %f894, %f895, %f921, %f1111, %f1111, %f924, %r62, %f363, %f926, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f263, [%rd2];
|
|
abs.f32 %f934, %f263;
|
|
ld.local.f32 %f264, [%rd2+4];
|
|
abs.f32 %f935, %f264;
|
|
add.f32 %f936, %f934, %f935;
|
|
ld.local.f32 %f265, [%rd2+8];
|
|
abs.f32 %f937, %f265;
|
|
add.f32 %f938, %f936, %f937;
|
|
setp.eq.f32 %p55, %f938, 0f00000000;
|
|
@%p55 bra BB0_72;
|
|
|
|
mul.f32 %f939, %f264, %f264;
|
|
fma.rn.f32 %f940, %f263, %f263, %f939;
|
|
fma.rn.f32 %f941, %f265, %f265, %f940;
|
|
sqrt.rn.f32 %f942, %f941;
|
|
max.f32 %f266, %f942, %f4;
|
|
add.f32 %f943, %f5, %f266;
|
|
fma.rn.f32 %f267, %f943, 0fBF13CD3A, %f1138;
|
|
fma.rn.f32 %f268, %f943, 0fBF13CD3A, %f1139;
|
|
fma.rn.f32 %f269, %f943, 0f3F13CD3A, %f1140;
|
|
setp.geu.f32 %p56, %f266, %f1167;
|
|
@%p56 bra BB0_72;
|
|
|
|
mov.f32 %f1164, %f267;
|
|
mov.f32 %f1165, %f268;
|
|
mov.f32 %f1166, %f269;
|
|
mov.f32 %f1167, %f266;
|
|
|
|
BB0_72:
|
|
sub.f32 %f1114, %f1140, %f76;
|
|
mov.f32 %f1113, 0f3F13CD3A;
|
|
mov.f32 %f1112, 0fBF13CD3A;
|
|
ld.global.f32 %f952, [voxelSize];
|
|
ld.global.f32 %f953, [voxelSize+4];
|
|
mul.f32 %f954, %f953, 0fBF13CD3A;
|
|
fma.rn.f32 %f955, %f952, 0f3F13CD3A, %f954;
|
|
ld.global.f32 %f956, [voxelSize+8];
|
|
fma.rn.f32 %f957, %f956, 0f3F13CD3A, %f955;
|
|
abs.f32 %f958, %f957;
|
|
mul.f32 %f951, %f958, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r106, [root];
|
|
sub.f32 %f944, %f1138, %f76;
|
|
// inline asm
|
|
call _rt_trace_64, (%r106, %f944, %f895, %f1114, %f1113, %f1112, %f1113, %r62, %f363, %f951, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f275, [%rd2];
|
|
abs.f32 %f959, %f275;
|
|
ld.local.f32 %f276, [%rd2+4];
|
|
abs.f32 %f960, %f276;
|
|
add.f32 %f961, %f959, %f960;
|
|
ld.local.f32 %f277, [%rd2+8];
|
|
abs.f32 %f962, %f277;
|
|
add.f32 %f963, %f961, %f962;
|
|
setp.eq.f32 %p57, %f963, 0f00000000;
|
|
@%p57 bra BB0_75;
|
|
|
|
mul.f32 %f964, %f276, %f276;
|
|
fma.rn.f32 %f965, %f275, %f275, %f964;
|
|
fma.rn.f32 %f966, %f277, %f277, %f965;
|
|
sqrt.rn.f32 %f967, %f966;
|
|
max.f32 %f278, %f967, %f4;
|
|
add.f32 %f968, %f5, %f278;
|
|
fma.rn.f32 %f279, %f968, 0f3F13CD3A, %f1138;
|
|
fma.rn.f32 %f280, %f968, 0fBF13CD3A, %f1139;
|
|
fma.rn.f32 %f281, %f968, 0f3F13CD3A, %f1140;
|
|
setp.geu.f32 %p58, %f278, %f1167;
|
|
@%p58 bra BB0_75;
|
|
|
|
mov.f32 %f1164, %f279;
|
|
mov.f32 %f1165, %f280;
|
|
mov.f32 %f1166, %f281;
|
|
mov.f32 %f1167, %f278;
|
|
|
|
BB0_75:
|
|
sub.f32 %f1099, %f1138, %f76;
|
|
mov.f32 %f1098, 0f3F13CD3A;
|
|
mov.f32 %f1097, 0fBF13CD3A;
|
|
ld.global.f32 %f977, [voxelSize];
|
|
ld.global.f32 %f978, [voxelSize+4];
|
|
mul.f32 %f979, %f978, 0fBF13CD3A;
|
|
fma.rn.f32 %f980, %f977, 0f3F13CD3A, %f979;
|
|
ld.global.f32 %f981, [voxelSize+8];
|
|
fma.rn.f32 %f982, %f981, 0fBF13CD3A, %f980;
|
|
abs.f32 %f983, %f982;
|
|
mul.f32 %f976, %f983, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r109, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r109, %f1099, %f895, %f896, %f1098, %f1097, %f1097, %r62, %f363, %f976, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f286, [%rd2];
|
|
abs.f32 %f984, %f286;
|
|
ld.local.f32 %f287, [%rd2+4];
|
|
abs.f32 %f985, %f287;
|
|
add.f32 %f986, %f984, %f985;
|
|
ld.local.f32 %f288, [%rd2+8];
|
|
abs.f32 %f987, %f288;
|
|
add.f32 %f988, %f986, %f987;
|
|
setp.eq.f32 %p59, %f988, 0f00000000;
|
|
@%p59 bra BB0_78;
|
|
|
|
mul.f32 %f989, %f287, %f287;
|
|
fma.rn.f32 %f990, %f286, %f286, %f989;
|
|
fma.rn.f32 %f991, %f288, %f288, %f990;
|
|
sqrt.rn.f32 %f992, %f991;
|
|
max.f32 %f289, %f992, %f4;
|
|
add.f32 %f993, %f5, %f289;
|
|
fma.rn.f32 %f290, %f993, 0f3F13CD3A, %f1138;
|
|
fma.rn.f32 %f291, %f993, 0fBF13CD3A, %f1139;
|
|
fma.rn.f32 %f292, %f993, 0fBF13CD3A, %f1140;
|
|
setp.geu.f32 %p60, %f289, %f1167;
|
|
@%p60 bra BB0_78;
|
|
|
|
mov.f32 %f1164, %f290;
|
|
mov.f32 %f1165, %f291;
|
|
mov.f32 %f1166, %f292;
|
|
mov.f32 %f1167, %f289;
|
|
|
|
BB0_78:
|
|
mov.f32 %f1101, 0f3F13CD3A;
|
|
mov.f32 %f1100, 0fBF13CD3A;
|
|
ld.global.f32 %f1002, [voxelSize];
|
|
ld.global.f32 %f1003, [voxelSize+4];
|
|
mul.f32 %f1004, %f1003, 0f3F13CD3A;
|
|
fma.rn.f32 %f1005, %f1002, 0fBF13CD3A, %f1004;
|
|
ld.global.f32 %f1006, [voxelSize+8];
|
|
fma.rn.f32 %f1007, %f1006, 0fBF13CD3A, %f1005;
|
|
abs.f32 %f1008, %f1007;
|
|
mul.f32 %f1001, %f1008, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r112, [root];
|
|
sub.f32 %f995, %f1139, %f76;
|
|
// inline asm
|
|
call _rt_trace_64, (%r112, %f894, %f995, %f896, %f1100, %f1101, %f1100, %r62, %f363, %f1001, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f298, [%rd2];
|
|
abs.f32 %f1009, %f298;
|
|
ld.local.f32 %f299, [%rd2+4];
|
|
abs.f32 %f1010, %f299;
|
|
add.f32 %f1011, %f1009, %f1010;
|
|
ld.local.f32 %f300, [%rd2+8];
|
|
abs.f32 %f1012, %f300;
|
|
add.f32 %f1013, %f1011, %f1012;
|
|
setp.eq.f32 %p61, %f1013, 0f00000000;
|
|
@%p61 bra BB0_81;
|
|
|
|
mul.f32 %f1014, %f299, %f299;
|
|
fma.rn.f32 %f1015, %f298, %f298, %f1014;
|
|
fma.rn.f32 %f1016, %f300, %f300, %f1015;
|
|
sqrt.rn.f32 %f1017, %f1016;
|
|
max.f32 %f301, %f1017, %f4;
|
|
add.f32 %f1018, %f5, %f301;
|
|
fma.rn.f32 %f302, %f1018, 0fBF13CD3A, %f1138;
|
|
fma.rn.f32 %f303, %f1018, 0f3F13CD3A, %f1139;
|
|
fma.rn.f32 %f304, %f1018, 0fBF13CD3A, %f1140;
|
|
setp.geu.f32 %p62, %f301, %f1167;
|
|
@%p62 bra BB0_81;
|
|
|
|
mov.f32 %f1164, %f302;
|
|
mov.f32 %f1165, %f303;
|
|
mov.f32 %f1166, %f304;
|
|
mov.f32 %f1167, %f301;
|
|
|
|
BB0_81:
|
|
sub.f32 %f1115, %f1139, %f76;
|
|
sub.f32 %f1104, %f1140, %f76;
|
|
mov.f32 %f1103, 0f3F13CD3A;
|
|
mov.f32 %f1102, 0fBF13CD3A;
|
|
ld.global.f32 %f1027, [voxelSize];
|
|
ld.global.f32 %f1028, [voxelSize+4];
|
|
mul.f32 %f1029, %f1028, 0f3F13CD3A;
|
|
fma.rn.f32 %f1030, %f1027, 0fBF13CD3A, %f1029;
|
|
ld.global.f32 %f1031, [voxelSize+8];
|
|
fma.rn.f32 %f1032, %f1031, 0f3F13CD3A, %f1030;
|
|
abs.f32 %f1033, %f1032;
|
|
mul.f32 %f1026, %f1033, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r115, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r115, %f894, %f1115, %f1104, %f1102, %f1103, %f1103, %r62, %f363, %f1026, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f309, [%rd2];
|
|
abs.f32 %f1034, %f309;
|
|
ld.local.f32 %f310, [%rd2+4];
|
|
abs.f32 %f1035, %f310;
|
|
add.f32 %f1036, %f1034, %f1035;
|
|
ld.local.f32 %f311, [%rd2+8];
|
|
abs.f32 %f1037, %f311;
|
|
add.f32 %f1038, %f1036, %f1037;
|
|
setp.eq.f32 %p63, %f1038, 0f00000000;
|
|
@%p63 bra BB0_84;
|
|
|
|
mul.f32 %f1039, %f310, %f310;
|
|
fma.rn.f32 %f1040, %f309, %f309, %f1039;
|
|
fma.rn.f32 %f1041, %f311, %f311, %f1040;
|
|
sqrt.rn.f32 %f1042, %f1041;
|
|
max.f32 %f312, %f1042, %f4;
|
|
add.f32 %f1043, %f5, %f312;
|
|
fma.rn.f32 %f313, %f1043, 0fBF13CD3A, %f1138;
|
|
fma.rn.f32 %f314, %f1043, 0f3F13CD3A, %f1139;
|
|
fma.rn.f32 %f315, %f1043, 0f3F13CD3A, %f1140;
|
|
setp.geu.f32 %p64, %f312, %f1167;
|
|
@%p64 bra BB0_84;
|
|
|
|
mov.f32 %f1164, %f313;
|
|
mov.f32 %f1165, %f314;
|
|
mov.f32 %f1166, %f315;
|
|
mov.f32 %f1167, %f312;
|
|
|
|
BB0_84:
|
|
sub.f32 %f1116, %f1139, %f76;
|
|
sub.f32 %f1107, %f1138, %f76;
|
|
sub.f32 %f1106, %f1140, %f76;
|
|
mov.f32 %f1105, 0f3F13CD3A;
|
|
ld.global.f32 %f1052, [voxelSize];
|
|
ld.global.f32 %f1053, [voxelSize+4];
|
|
mul.f32 %f1054, %f1053, 0f3F13CD3A;
|
|
fma.rn.f32 %f1055, %f1052, 0f3F13CD3A, %f1054;
|
|
ld.global.f32 %f1056, [voxelSize+8];
|
|
fma.rn.f32 %f1057, %f1056, 0f3F13CD3A, %f1055;
|
|
abs.f32 %f1058, %f1057;
|
|
mul.f32 %f1051, %f1058, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r118, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r118, %f1107, %f1116, %f1106, %f1105, %f1105, %f1105, %r62, %f363, %f1051, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f320, [%rd2];
|
|
abs.f32 %f1059, %f320;
|
|
ld.local.f32 %f321, [%rd2+4];
|
|
abs.f32 %f1060, %f321;
|
|
add.f32 %f1061, %f1059, %f1060;
|
|
ld.local.f32 %f322, [%rd2+8];
|
|
abs.f32 %f1062, %f322;
|
|
add.f32 %f1063, %f1061, %f1062;
|
|
setp.eq.f32 %p65, %f1063, 0f00000000;
|
|
@%p65 bra BB0_87;
|
|
|
|
mul.f32 %f1064, %f321, %f321;
|
|
fma.rn.f32 %f1065, %f320, %f320, %f1064;
|
|
fma.rn.f32 %f1066, %f322, %f322, %f1065;
|
|
sqrt.rn.f32 %f1067, %f1066;
|
|
max.f32 %f323, %f1067, %f4;
|
|
add.f32 %f1068, %f5, %f323;
|
|
fma.rn.f32 %f324, %f1068, 0f3F13CD3A, %f1138;
|
|
fma.rn.f32 %f325, %f1068, 0f3F13CD3A, %f1139;
|
|
fma.rn.f32 %f326, %f1068, 0f3F13CD3A, %f1140;
|
|
setp.geu.f32 %p66, %f323, %f1167;
|
|
@%p66 bra BB0_87;
|
|
|
|
mov.f32 %f1164, %f324;
|
|
mov.f32 %f1165, %f325;
|
|
mov.f32 %f1166, %f326;
|
|
mov.f32 %f1167, %f323;
|
|
|
|
BB0_87:
|
|
sub.f32 %f1117, %f1139, %f76;
|
|
sub.f32 %f1110, %f1138, %f76;
|
|
mov.f32 %f1109, 0f3F13CD3A;
|
|
mov.f32 %f1108, 0fBF13CD3A;
|
|
ld.global.f32 %f1077, [voxelSize];
|
|
ld.global.f32 %f1078, [voxelSize+4];
|
|
mul.f32 %f1079, %f1078, 0f3F13CD3A;
|
|
fma.rn.f32 %f1080, %f1077, 0f3F13CD3A, %f1079;
|
|
ld.global.f32 %f1081, [voxelSize+8];
|
|
fma.rn.f32 %f1082, %f1081, 0fBF13CD3A, %f1080;
|
|
abs.f32 %f1083, %f1082;
|
|
mul.f32 %f1076, %f1083, 0f3FDDB3D7;
|
|
st.local.u32 [%rd2+8], %r62;
|
|
st.local.u32 [%rd2+4], %r62;
|
|
st.local.u32 [%rd2], %r62;
|
|
ld.global.u32 %r121, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r121, %f1110, %f1117, %f896, %f1109, %f1109, %f1108, %r62, %f363, %f1076, %rd42, %r29);
|
|
// inline asm
|
|
ld.local.f32 %f331, [%rd2];
|
|
abs.f32 %f1084, %f331;
|
|
ld.local.f32 %f332, [%rd2+4];
|
|
abs.f32 %f1085, %f332;
|
|
add.f32 %f1086, %f1084, %f1085;
|
|
ld.local.f32 %f333, [%rd2+8];
|
|
abs.f32 %f1087, %f333;
|
|
add.f32 %f1088, %f1086, %f1087;
|
|
setp.eq.f32 %p67, %f1088, 0f00000000;
|
|
@%p67 bra BB0_88;
|
|
|
|
mul.f32 %f1089, %f332, %f332;
|
|
fma.rn.f32 %f1090, %f331, %f331, %f1089;
|
|
fma.rn.f32 %f1091, %f333, %f333, %f1090;
|
|
sqrt.rn.f32 %f1092, %f1091;
|
|
max.f32 %f1093, %f1092, %f4;
|
|
add.f32 %f1094, %f5, %f1093;
|
|
fma.rn.f32 %f334, %f1094, 0f3F13CD3A, %f1138;
|
|
fma.rn.f32 %f335, %f1094, 0f3F13CD3A, %f1139;
|
|
fma.rn.f32 %f336, %f1094, 0fBF13CD3A, %f1140;
|
|
setp.geu.f32 %p68, %f1093, %f1167;
|
|
mov.f32 %f1138, %f1164;
|
|
mov.f32 %f1139, %f1165;
|
|
mov.f32 %f1140, %f1166;
|
|
@%p68 bra BB0_91;
|
|
|
|
mov.f32 %f1138, %f334;
|
|
mov.f32 %f1139, %f335;
|
|
mov.f32 %f1140, %f336;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_88:
|
|
mov.f32 %f1138, %f1164;
|
|
mov.f32 %f1139, %f1165;
|
|
mov.f32 %f1140, %f1166;
|
|
|
|
BB0_91:
|
|
add.s32 %r136, %r136, 1;
|
|
setp.lt.s32 %p69, %r136, 2;
|
|
@%p69 bra BB0_23;
|
|
|
|
mov.u64 %rd84, 0;
|
|
mov.u32 %r128, 2;
|
|
cvt.u64.u32 %rd67, %r2;
|
|
cvt.u64.u32 %rd68, %r3;
|
|
mov.u64 %rd71, image_HDR;
|
|
cvta.global.u64 %rd66, %rd71;
|
|
mov.u32 %r125, 16;
|
|
// inline asm
|
|
call (%rd65), _rt_buffer_get_64, (%rd66, %r128, %r125, %rd67, %rd68, %rd84, %rd84);
|
|
// inline asm
|
|
st.v4.f32 [%rd65], {%f1138, %f1139, %f1140, %f402};
|
|
|
|
BB0_94:
|
|
ret;
|
|
}
|
|
|
|
|