1524 lines
44 KiB
Plaintext
1524 lines
44 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 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Dir[1];
|
|
.global .align 8 .b8 texCoords[8];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 1 .b8 lightmapDirect[1];
|
|
.global .align 1 .b8 lightmapDirectDir[1];
|
|
.global .texref albedoTex;
|
|
.global .align 4 .u32 samples;
|
|
.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[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 8 .b8 _ZN21rti_internal_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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 4 .u32 _ZN21rti_internal_typeenum9texCoordsE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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 16 .b8 _ZN21rti_internal_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation9texCoordsE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[40];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<100>;
|
|
.reg .b16 %rs<44>;
|
|
.reg .f32 %f<715>;
|
|
.reg .b32 %r<332>;
|
|
.reg .b64 %rd<143>;
|
|
|
|
|
|
mov.u64 %rd142, __local_depot0;
|
|
cvta.local.u64 %SP, %rd142;
|
|
ld.global.u32 %r1, [samples];
|
|
ld.global.v2.u32 {%r100, %r101}, [pixelID];
|
|
cvt.u64.u32 %rd20, %r100;
|
|
cvt.u64.u32 %rd21, %r101;
|
|
mov.u64 %rd24, uvnormal;
|
|
cvta.global.u64 %rd19, %rd24;
|
|
mov.u32 %r98, 2;
|
|
mov.u32 %r99, 4;
|
|
mov.u64 %rd23, 0;
|
|
// inline asm
|
|
call (%rd18), _rt_buffer_get_64, (%rd19, %r98, %r99, %rd20, %rd21, %rd23, %rd23);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd18];
|
|
shr.u32 %r104, %r2, 16;
|
|
cvt.u16.u32 %rs1, %r104;
|
|
and.b16 %rs2, %rs1, 255;
|
|
cvt.u16.u32 %rs3, %r2;
|
|
or.b16 %rs4, %rs3, %rs2;
|
|
setp.eq.s16 %p4, %rs4, 0;
|
|
mov.f32 %f670, 0f00000000;
|
|
mov.f32 %f671, %f670;
|
|
mov.f32 %f672, %f670;
|
|
@%p4 bra BB0_2;
|
|
|
|
ld.u8 %rs5, [%rd18+1];
|
|
and.b16 %rs7, %rs3, 255;
|
|
cvt.rn.f32.u16 %f145, %rs7;
|
|
div.rn.f32 %f146, %f145, 0f437F0000;
|
|
fma.rn.f32 %f147, %f146, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f148, %rs5;
|
|
div.rn.f32 %f149, %f148, 0f437F0000;
|
|
fma.rn.f32 %f150, %f149, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f151, %rs2;
|
|
div.rn.f32 %f152, %f151, 0f437F0000;
|
|
fma.rn.f32 %f153, %f152, 0f40000000, 0fBF800000;
|
|
mul.f32 %f154, %f150, %f150;
|
|
fma.rn.f32 %f155, %f147, %f147, %f154;
|
|
fma.rn.f32 %f156, %f153, %f153, %f155;
|
|
sqrt.rn.f32 %f157, %f156;
|
|
rcp.rn.f32 %f158, %f157;
|
|
mul.f32 %f670, %f147, %f158;
|
|
mul.f32 %f671, %f150, %f158;
|
|
mul.f32 %f672, %f153, %f158;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
ld.global.v2.u32 {%r108, %r109}, [tileInfo];
|
|
add.s32 %r3, %r105, %r108;
|
|
add.s32 %r4, %r106, %r109;
|
|
setp.eq.f32 %p5, %f671, 0f00000000;
|
|
setp.eq.f32 %p6, %f670, 0f00000000;
|
|
and.pred %p7, %p6, %p5;
|
|
setp.eq.f32 %p8, %f672, 0f00000000;
|
|
and.pred %p9, %p7, %p8;
|
|
@%p9 bra BB0_101;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_101:
|
|
ld.global.u32 %r330, [imageEnabled];
|
|
and.b32 %r280, %r330, 1;
|
|
setp.eq.b32 %p96, %r280, 1;
|
|
@!%p96 bra BB0_103;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_102:
|
|
cvt.u64.u32 %rd106, %r3;
|
|
cvt.u64.u32 %rd107, %r4;
|
|
mov.u64 %rd110, image;
|
|
cvta.global.u64 %rd105, %rd110;
|
|
mov.u64 %rd109, 0;
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd105, %r98, %r99, %rd106, %rd107, %rd109, %rd109);
|
|
// inline asm
|
|
mov.u16 %rs38, 0;
|
|
st.v4.u8 [%rd104], {%rs38, %rs38, %rs38, %rs38};
|
|
ld.global.u32 %r330, [imageEnabled];
|
|
|
|
BB0_103:
|
|
and.b32 %r283, %r330, 4;
|
|
setp.eq.s32 %p97, %r283, 0;
|
|
@%p97 bra BB0_105;
|
|
|
|
cvt.u64.u32 %rd113, %r3;
|
|
cvt.u64.u32 %rd114, %r4;
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r285, 8;
|
|
mov.u64 %rd116, 0;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r98, %r285, %rd113, %rd114, %rd116, %rd116);
|
|
// inline asm
|
|
mov.f32 %f619, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f619;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs40, 0;
|
|
st.v4.u16 [%rd111], {%rs39, %rs39, %rs39, %rs40};
|
|
ld.global.u32 %r330, [imageEnabled];
|
|
|
|
BB0_105:
|
|
and.b32 %r286, %r330, 16;
|
|
setp.eq.s32 %p98, %r286, 0;
|
|
@%p98 bra BB0_107;
|
|
|
|
cvt.u64.u32 %rd120, %r3;
|
|
cvt.u64.u32 %rd121, %r4;
|
|
mov.u64 %rd124, image_HDR2;
|
|
cvta.global.u64 %rd119, %rd124;
|
|
mov.u32 %r288, 8;
|
|
mov.u64 %rd123, 0;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r98, %r288, %rd120, %rd121, %rd123, %rd123);
|
|
// inline asm
|
|
mov.f32 %f620, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f620;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs42, 0;
|
|
st.v4.u16 [%rd118], {%rs41, %rs41, %rs41, %rs42};
|
|
ld.global.u32 %r330, [imageEnabled];
|
|
|
|
BB0_107:
|
|
and.b32 %r289, %r330, 64;
|
|
setp.eq.s32 %p99, %r289, 0;
|
|
@%p99 bra BB0_109;
|
|
|
|
cvt.u64.u32 %rd127, %r3;
|
|
cvt.u64.u32 %rd128, %r4;
|
|
mov.u64 %rd131, image_Dir;
|
|
cvta.global.u64 %rd126, %rd131;
|
|
mov.u64 %rd130, 0;
|
|
// inline asm
|
|
call (%rd125), _rt_buffer_get_64, (%rd126, %r98, %r99, %rd127, %rd128, %rd130, %rd130);
|
|
// inline asm
|
|
mov.u16 %rs43, 0;
|
|
st.v4.u8 [%rd125], {%rs43, %rs43, %rs43, %rs43};
|
|
bra.uni BB0_109;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r119, %r120}, [pixelID];
|
|
cvt.u64.u32 %rd27, %r119;
|
|
cvt.u64.u32 %rd28, %r120;
|
|
mov.u64 %rd43, lightmapDirect;
|
|
cvta.global.u64 %rd26, %rd43;
|
|
mov.u32 %r114, 8;
|
|
// inline asm
|
|
call (%rd25), _rt_buffer_get_64, (%rd26, %r98, %r114, %rd27, %rd28, %rd23, %rd23);
|
|
// inline asm
|
|
ld.v4.u16 {%rs12, %rs13, %rs14, %rs15}, [%rd25];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f159, %rs12;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f160, %rs13;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f161, %rs14;}
|
|
|
|
// inline asm
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd33, %r123;
|
|
cvt.u64.u32 %rd34, %r124;
|
|
mov.u64 %rd44, uvpos;
|
|
cvta.global.u64 %rd32, %rd44;
|
|
mov.u32 %r116, 12;
|
|
// inline asm
|
|
call (%rd31), _rt_buffer_get_64, (%rd32, %r98, %r116, %rd33, %rd34, %rd23, %rd23);
|
|
// inline asm
|
|
ld.f32 %f168, [%rd31+8];
|
|
ld.f32 %f169, [%rd31+4];
|
|
ld.f32 %f170, [%rd31];
|
|
mul.f32 %f171, %f170, 0f3456BF95;
|
|
mul.f32 %f172, %f169, 0f3456BF95;
|
|
mul.f32 %f173, %f168, 0f3456BF95;
|
|
abs.f32 %f174, %f670;
|
|
div.rn.f32 %f175, %f171, %f174;
|
|
abs.f32 %f176, %f671;
|
|
div.rn.f32 %f177, %f172, %f176;
|
|
abs.f32 %f178, %f672;
|
|
div.rn.f32 %f179, %f173, %f178;
|
|
abs.f32 %f180, %f175;
|
|
abs.f32 %f181, %f177;
|
|
abs.f32 %f182, %f179;
|
|
mov.f32 %f183, 0f38D1B717;
|
|
max.f32 %f184, %f180, %f183;
|
|
max.f32 %f185, %f181, %f183;
|
|
max.f32 %f186, %f182, %f183;
|
|
fma.rn.f32 %f10, %f670, %f184, %f170;
|
|
fma.rn.f32 %f11, %f671, %f185, %f169;
|
|
fma.rn.f32 %f12, %f672, %f186, %f168;
|
|
setp.gt.f32 %p10, %f174, %f178;
|
|
neg.f32 %f187, %f671;
|
|
selp.f32 %f188, %f187, 0f00000000, %p10;
|
|
neg.f32 %f189, %f672;
|
|
selp.f32 %f190, %f670, %f189, %p10;
|
|
selp.f32 %f191, 0f00000000, %f671, %p10;
|
|
mul.f32 %f192, %f190, %f190;
|
|
fma.rn.f32 %f193, %f188, %f188, %f192;
|
|
fma.rn.f32 %f194, %f191, %f191, %f193;
|
|
sqrt.rn.f32 %f195, %f194;
|
|
rcp.rn.f32 %f196, %f195;
|
|
mul.f32 %f13, %f188, %f196;
|
|
mul.f32 %f14, %f190, %f196;
|
|
mul.f32 %f15, %f191, %f196;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd39, %r127;
|
|
cvt.u64.u32 %rd40, %r128;
|
|
mov.u64 %rd45, rnd_seeds;
|
|
cvta.global.u64 %rd38, %rd45;
|
|
// inline asm
|
|
call (%rd37), _rt_buffer_get_64, (%rd38, %r98, %r99, %rd39, %rd40, %rd23, %rd23);
|
|
// inline asm
|
|
mov.f32 %f679, 0f00000000;
|
|
setp.lt.s32 %p11, %r1, 1;
|
|
mov.f32 %f680, %f679;
|
|
mov.f32 %f681, %f679;
|
|
mov.f32 %f682, %f679;
|
|
mov.f32 %f683, %f679;
|
|
mov.f32 %f684, %f679;
|
|
@%p11 bra BB0_54;
|
|
|
|
cvt.rn.f32.s32 %f203, %r1;
|
|
rcp.rn.f32 %f16, %f203;
|
|
ld.u32 %r305, [%rd37];
|
|
mul.f32 %f17, %f10, 0f3456BF95;
|
|
mul.f32 %f18, %f11, 0f3456BF95;
|
|
mul.f32 %f19, %f12, 0f3456BF95;
|
|
mul.f32 %f204, %f670, %f14;
|
|
mul.f32 %f205, %f671, %f13;
|
|
sub.f32 %f20, %f205, %f204;
|
|
mul.f32 %f206, %f672, %f13;
|
|
mul.f32 %f207, %f670, %f15;
|
|
sub.f32 %f21, %f207, %f206;
|
|
mul.f32 %f208, %f671, %f15;
|
|
mul.f32 %f209, %f672, %f14;
|
|
sub.f32 %f22, %f209, %f208;
|
|
mov.f32 %f679, 0f00000000;
|
|
mov.u32 %r131, 0;
|
|
abs.f32 %f210, %f18;
|
|
abs.f32 %f211, %f17;
|
|
max.f32 %f212, %f211, %f210;
|
|
abs.f32 %f213, %f19;
|
|
max.f32 %f214, %f212, %f213;
|
|
mov.u32 %r302, %r131;
|
|
mov.f32 %f680, %f679;
|
|
mov.f32 %f681, %f679;
|
|
mov.f32 %f682, %f679;
|
|
mov.f32 %f683, %f679;
|
|
mov.f32 %f684, %f679;
|
|
|
|
BB0_5:
|
|
max.f32 %f30, %f214, %f183;
|
|
mov.u32 %r304, %r131;
|
|
|
|
BB0_6:
|
|
cvt.rn.f32.s32 %f667, %r302;
|
|
mad.lo.s32 %r133, %r305, 1664525, 1013904223;
|
|
and.b32 %r134, %r133, 16777215;
|
|
cvt.rn.f32.u32 %f216, %r134;
|
|
fma.rn.f32 %f217, %f216, 0f33800000, %f667;
|
|
mul.f32 %f37, %f16, %f217;
|
|
mad.lo.s32 %r305, %r133, 1664525, 1013904223;
|
|
and.b32 %r135, %r305, 16777215;
|
|
cvt.rn.f32.u32 %f218, %r135;
|
|
cvt.rn.f32.s32 %f219, %r304;
|
|
fma.rn.f32 %f220, %f218, 0f33800000, %f219;
|
|
mul.f32 %f221, %f16, %f220;
|
|
mul.f32 %f222, %f37, %f37;
|
|
mov.f32 %f223, 0f3F800000;
|
|
sub.f32 %f224, %f223, %f222;
|
|
mov.f32 %f225, 0f00000000;
|
|
max.f32 %f226, %f225, %f224;
|
|
sqrt.rn.f32 %f38, %f226;
|
|
mul.f32 %f691, %f221, 0f40C90FDB;
|
|
abs.f32 %f40, %f691;
|
|
setp.neu.f32 %p12, %f40, 0f7F800000;
|
|
mov.f32 %f685, %f691;
|
|
@%p12 bra BB0_8;
|
|
|
|
mul.rn.f32 %f685, %f691, %f225;
|
|
|
|
BB0_8:
|
|
mul.f32 %f228, %f685, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r315, %f228;
|
|
cvt.rn.f32.s32 %f229, %r315;
|
|
neg.f32 %f230, %f229;
|
|
mov.f32 %f231, 0f3FC90FDA;
|
|
fma.rn.f32 %f232, %f230, %f231, %f685;
|
|
mov.f32 %f233, 0f33A22168;
|
|
fma.rn.f32 %f234, %f230, %f233, %f232;
|
|
mov.f32 %f235, 0f27C234C5;
|
|
fma.rn.f32 %f686, %f230, %f235, %f234;
|
|
abs.f32 %f236, %f685;
|
|
setp.leu.f32 %p13, %f236, 0f47CE4780;
|
|
@%p13 bra BB0_19;
|
|
|
|
add.u64 %rd47, %SP, 12;
|
|
cvta.to.local.u64 %rd3, %rd47;
|
|
mov.b32 %r12, %f685;
|
|
shr.u32 %r13, %r12, 23;
|
|
shl.b32 %r138, %r12, 8;
|
|
or.b32 %r14, %r138, -2147483648;
|
|
mov.u32 %r306, 0;
|
|
mov.u64 %rd139, 0;
|
|
mov.u64 %rd138, %rd3;
|
|
mov.u32 %r307, %r306;
|
|
|
|
BB0_10:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd48, %rd139, 2;
|
|
mov.u64 %rd49, __cudart_i2opi_f;
|
|
add.s64 %rd50, %rd49, %rd48;
|
|
ld.const.u32 %r141, [%rd50];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r139, %r141, %r14, %r307;
|
|
madc.hi.u32 %r307, %r141, %r14, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd138], %r139;
|
|
add.s32 %r306, %r306, 1;
|
|
cvt.s64.s32 %rd139, %r306;
|
|
mul.wide.s32 %rd53, %r306, 4;
|
|
add.s64 %rd138, %rd3, %rd53;
|
|
setp.ne.s32 %p14, %r306, 6;
|
|
@%p14 bra BB0_10;
|
|
|
|
add.u64 %rd137, %SP, 12;
|
|
and.b32 %r144, %r13, 255;
|
|
add.s32 %r145, %r144, -128;
|
|
shr.u32 %r146, %r145, 5;
|
|
and.b32 %r19, %r12, -2147483648;
|
|
cvta.to.local.u64 %rd55, %rd137;
|
|
st.local.u32 [%rd55+24], %r307;
|
|
mov.u32 %r147, 6;
|
|
sub.s32 %r148, %r147, %r146;
|
|
mul.wide.s32 %rd56, %r148, 4;
|
|
add.s64 %rd8, %rd55, %rd56;
|
|
ld.local.u32 %r308, [%rd8];
|
|
ld.local.u32 %r309, [%rd8+-4];
|
|
and.b32 %r22, %r13, 31;
|
|
setp.eq.s32 %p15, %r22, 0;
|
|
@%p15 bra BB0_13;
|
|
|
|
mov.u32 %r149, 32;
|
|
sub.s32 %r150, %r149, %r22;
|
|
shr.u32 %r151, %r309, %r150;
|
|
shl.b32 %r152, %r308, %r22;
|
|
add.s32 %r308, %r151, %r152;
|
|
ld.local.u32 %r153, [%rd8+-8];
|
|
shr.u32 %r154, %r153, %r150;
|
|
shl.b32 %r155, %r309, %r22;
|
|
add.s32 %r309, %r154, %r155;
|
|
|
|
BB0_13:
|
|
shr.u32 %r156, %r309, 30;
|
|
shl.b32 %r157, %r308, 2;
|
|
add.s32 %r310, %r156, %r157;
|
|
shl.b32 %r28, %r309, 2;
|
|
shr.u32 %r158, %r310, 31;
|
|
shr.u32 %r159, %r308, 30;
|
|
add.s32 %r29, %r158, %r159;
|
|
setp.eq.s32 %p16, %r158, 0;
|
|
@%p16 bra BB0_14;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_14:
|
|
mov.u32 %r311, %r19;
|
|
mov.u32 %r312, %r28;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
not.b32 %r160, %r310;
|
|
neg.s32 %r312, %r28;
|
|
setp.eq.s32 %p17, %r28, 0;
|
|
selp.u32 %r161, 1, 0, %p17;
|
|
add.s32 %r310, %r161, %r160;
|
|
xor.b32 %r311, %r19, -2147483648;
|
|
|
|
BB0_16:
|
|
clz.b32 %r314, %r310;
|
|
setp.eq.s32 %p18, %r314, 0;
|
|
shl.b32 %r162, %r310, %r314;
|
|
mov.u32 %r163, 32;
|
|
sub.s32 %r164, %r163, %r314;
|
|
shr.u32 %r165, %r312, %r164;
|
|
add.s32 %r166, %r165, %r162;
|
|
selp.b32 %r37, %r310, %r166, %p18;
|
|
mov.u32 %r167, -921707870;
|
|
mul.hi.u32 %r313, %r37, %r167;
|
|
setp.eq.s32 %p19, %r19, 0;
|
|
neg.s32 %r168, %r29;
|
|
selp.b32 %r315, %r29, %r168, %p19;
|
|
setp.lt.s32 %p20, %r313, 1;
|
|
@%p20 bra BB0_18;
|
|
|
|
mul.lo.s32 %r169, %r37, -921707870;
|
|
shr.u32 %r170, %r169, 31;
|
|
shl.b32 %r171, %r313, 1;
|
|
add.s32 %r313, %r170, %r171;
|
|
add.s32 %r314, %r314, 1;
|
|
|
|
BB0_18:
|
|
mov.u32 %r172, 126;
|
|
sub.s32 %r173, %r172, %r314;
|
|
shl.b32 %r174, %r173, 23;
|
|
add.s32 %r175, %r313, 1;
|
|
shr.u32 %r176, %r175, 7;
|
|
add.s32 %r177, %r176, 1;
|
|
shr.u32 %r178, %r177, 1;
|
|
add.s32 %r179, %r178, %r174;
|
|
or.b32 %r180, %r179, %r311;
|
|
mov.b32 %f686, %r180;
|
|
|
|
BB0_19:
|
|
mul.rn.f32 %f46, %f686, %f686;
|
|
add.s32 %r45, %r315, 1;
|
|
and.b32 %r46, %r45, 1;
|
|
setp.eq.s32 %p21, %r46, 0;
|
|
@%p21 bra BB0_21;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_21:
|
|
mov.f32 %f239, 0f3C08839E;
|
|
mov.f32 %f240, 0fB94CA1F9;
|
|
fma.rn.f32 %f687, %f240, %f46, %f239;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_20:
|
|
mov.f32 %f237, 0fBAB6061A;
|
|
mov.f32 %f238, 0f37CCF5CE;
|
|
fma.rn.f32 %f687, %f238, %f46, %f237;
|
|
|
|
BB0_22:
|
|
@%p21 bra BB0_24;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_24:
|
|
mov.f32 %f244, 0fBE2AAAA3;
|
|
fma.rn.f32 %f245, %f687, %f46, %f244;
|
|
fma.rn.f32 %f688, %f245, %f46, %f225;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_23:
|
|
mov.f32 %f241, 0f3D2AAAA5;
|
|
fma.rn.f32 %f242, %f687, %f46, %f241;
|
|
mov.f32 %f243, 0fBF000000;
|
|
fma.rn.f32 %f688, %f242, %f46, %f243;
|
|
|
|
BB0_25:
|
|
fma.rn.f32 %f689, %f688, %f686, %f686;
|
|
@%p21 bra BB0_27;
|
|
|
|
fma.rn.f32 %f689, %f688, %f46, %f223;
|
|
|
|
BB0_27:
|
|
and.b32 %r181, %r45, 2;
|
|
setp.eq.s32 %p24, %r181, 0;
|
|
@%p24 bra BB0_29;
|
|
|
|
mov.f32 %f249, 0fBF800000;
|
|
fma.rn.f32 %f689, %f689, %f249, %f225;
|
|
|
|
BB0_29:
|
|
@%p12 bra BB0_31;
|
|
|
|
mul.rn.f32 %f691, %f691, %f225;
|
|
|
|
BB0_31:
|
|
mov.f32 %f669, 0f27C234C5;
|
|
mov.f32 %f668, 0f33A22168;
|
|
mov.f32 %f666, 0f3FC90FDA;
|
|
mul.f32 %f251, %f691, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r325, %f251;
|
|
cvt.rn.f32.s32 %f252, %r325;
|
|
neg.f32 %f253, %f252;
|
|
fma.rn.f32 %f255, %f253, %f666, %f691;
|
|
fma.rn.f32 %f257, %f253, %f668, %f255;
|
|
fma.rn.f32 %f692, %f253, %f669, %f257;
|
|
abs.f32 %f259, %f691;
|
|
setp.leu.f32 %p26, %f259, 0f47CE4780;
|
|
@%p26 bra BB0_42;
|
|
|
|
add.u64 %rd58, %SP, 12;
|
|
cvta.to.local.u64 %rd9, %rd58;
|
|
mov.b32 %r48, %f691;
|
|
shr.u32 %r49, %r48, 23;
|
|
shl.b32 %r184, %r48, 8;
|
|
or.b32 %r50, %r184, -2147483648;
|
|
mov.u32 %r316, 0;
|
|
mov.u64 %rd140, %rd9;
|
|
mov.u64 %rd141, %rd23;
|
|
mov.u32 %r317, %r316;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd59, %rd141, 2;
|
|
mov.u64 %rd60, __cudart_i2opi_f;
|
|
add.s64 %rd61, %rd60, %rd59;
|
|
ld.const.u32 %r187, [%rd61];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r185, %r187, %r50, %r317;
|
|
madc.hi.u32 %r317, %r187, %r50, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd140], %r185;
|
|
add.s32 %r316, %r316, 1;
|
|
cvt.s64.s32 %rd141, %r316;
|
|
mul.wide.s32 %rd62, %r316, 4;
|
|
add.s64 %rd140, %rd9, %rd62;
|
|
setp.ne.s32 %p27, %r316, 6;
|
|
@%p27 bra BB0_33;
|
|
|
|
and.b32 %r190, %r49, 255;
|
|
add.s32 %r191, %r190, -128;
|
|
shr.u32 %r192, %r191, 5;
|
|
and.b32 %r55, %r48, -2147483648;
|
|
cvta.to.local.u64 %rd64, %rd58;
|
|
st.local.u32 [%rd64+24], %r317;
|
|
mov.u32 %r193, 6;
|
|
sub.s32 %r194, %r193, %r192;
|
|
mul.wide.s32 %rd65, %r194, 4;
|
|
add.s64 %rd15, %rd64, %rd65;
|
|
ld.local.u32 %r318, [%rd15];
|
|
ld.local.u32 %r319, [%rd15+-4];
|
|
and.b32 %r58, %r49, 31;
|
|
setp.eq.s32 %p28, %r58, 0;
|
|
@%p28 bra BB0_36;
|
|
|
|
mov.u32 %r195, 32;
|
|
sub.s32 %r196, %r195, %r58;
|
|
shr.u32 %r197, %r319, %r196;
|
|
shl.b32 %r198, %r318, %r58;
|
|
add.s32 %r318, %r197, %r198;
|
|
ld.local.u32 %r199, [%rd15+-8];
|
|
shr.u32 %r200, %r199, %r196;
|
|
shl.b32 %r201, %r319, %r58;
|
|
add.s32 %r319, %r200, %r201;
|
|
|
|
BB0_36:
|
|
shr.u32 %r202, %r319, 30;
|
|
shl.b32 %r203, %r318, 2;
|
|
add.s32 %r320, %r202, %r203;
|
|
shl.b32 %r64, %r319, 2;
|
|
shr.u32 %r204, %r320, 31;
|
|
shr.u32 %r205, %r318, 30;
|
|
add.s32 %r65, %r204, %r205;
|
|
setp.eq.s32 %p29, %r204, 0;
|
|
@%p29 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r321, %r55;
|
|
mov.u32 %r322, %r64;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r206, %r320;
|
|
neg.s32 %r322, %r64;
|
|
setp.eq.s32 %p30, %r64, 0;
|
|
selp.u32 %r207, 1, 0, %p30;
|
|
add.s32 %r320, %r207, %r206;
|
|
xor.b32 %r321, %r55, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r324, %r320;
|
|
setp.eq.s32 %p31, %r324, 0;
|
|
shl.b32 %r208, %r320, %r324;
|
|
mov.u32 %r209, 32;
|
|
sub.s32 %r210, %r209, %r324;
|
|
shr.u32 %r211, %r322, %r210;
|
|
add.s32 %r212, %r211, %r208;
|
|
selp.b32 %r73, %r320, %r212, %p31;
|
|
mov.u32 %r213, -921707870;
|
|
mul.hi.u32 %r323, %r73, %r213;
|
|
setp.eq.s32 %p32, %r55, 0;
|
|
neg.s32 %r214, %r65;
|
|
selp.b32 %r325, %r65, %r214, %p32;
|
|
setp.lt.s32 %p33, %r323, 1;
|
|
@%p33 bra BB0_41;
|
|
|
|
mul.lo.s32 %r215, %r73, -921707870;
|
|
shr.u32 %r216, %r215, 31;
|
|
shl.b32 %r217, %r323, 1;
|
|
add.s32 %r323, %r216, %r217;
|
|
add.s32 %r324, %r324, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r218, 126;
|
|
sub.s32 %r219, %r218, %r324;
|
|
shl.b32 %r220, %r219, 23;
|
|
add.s32 %r221, %r323, 1;
|
|
shr.u32 %r222, %r221, 7;
|
|
add.s32 %r223, %r222, 1;
|
|
shr.u32 %r224, %r223, 1;
|
|
add.s32 %r225, %r224, %r220;
|
|
or.b32 %r226, %r225, %r321;
|
|
mov.b32 %f692, %r226;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f63, %f692, %f692;
|
|
and.b32 %r81, %r325, 1;
|
|
setp.eq.s32 %p34, %r81, 0;
|
|
@%p34 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f262, 0f3C08839E;
|
|
mov.f32 %f263, 0fB94CA1F9;
|
|
fma.rn.f32 %f693, %f263, %f63, %f262;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f260, 0fBAB6061A;
|
|
mov.f32 %f261, 0f37CCF5CE;
|
|
fma.rn.f32 %f693, %f261, %f63, %f260;
|
|
|
|
BB0_45:
|
|
@%p34 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f267, 0fBE2AAAA3;
|
|
fma.rn.f32 %f268, %f693, %f63, %f267;
|
|
fma.rn.f32 %f694, %f268, %f63, %f225;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f264, 0f3D2AAAA5;
|
|
fma.rn.f32 %f265, %f693, %f63, %f264;
|
|
mov.f32 %f266, 0fBF000000;
|
|
fma.rn.f32 %f694, %f265, %f63, %f266;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f695, %f694, %f692, %f692;
|
|
@%p34 bra BB0_50;
|
|
|
|
fma.rn.f32 %f695, %f694, %f63, %f223;
|
|
|
|
BB0_50:
|
|
and.b32 %r227, %r325, 2;
|
|
setp.eq.s32 %p37, %r227, 0;
|
|
@%p37 bra BB0_52;
|
|
|
|
mov.f32 %f272, 0fBF800000;
|
|
fma.rn.f32 %f695, %f695, %f272, %f225;
|
|
|
|
BB0_52:
|
|
mul.f32 %f281, %f38, %f689;
|
|
add.u64 %rd66, %SP, 0;
|
|
cvta.to.local.u64 %rd67, %rd66;
|
|
mul.f32 %f282, %f38, %f695;
|
|
mul.f32 %f283, %f13, %f282;
|
|
mul.f32 %f284, %f14, %f282;
|
|
mul.f32 %f285, %f15, %f282;
|
|
fma.rn.f32 %f286, %f22, %f281, %f283;
|
|
fma.rn.f32 %f287, %f21, %f281, %f284;
|
|
fma.rn.f32 %f288, %f20, %f281, %f285;
|
|
fma.rn.f32 %f276, %f670, %f37, %f286;
|
|
fma.rn.f32 %f277, %f671, %f37, %f287;
|
|
fma.rn.f32 %f278, %f672, %f37, %f288;
|
|
mov.u32 %r229, 0;
|
|
st.local.u32 [%rd67+8], %r229;
|
|
st.local.u32 [%rd67+4], %r229;
|
|
st.local.u32 [%rd67], %r229;
|
|
ld.global.u32 %r228, [root];
|
|
mov.f32 %f280, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r228, %f10, %f11, %f12, %f276, %f277, %f278, %r229, %f30, %f280, %rd66, %r116);
|
|
// inline asm
|
|
ld.local.f32 %f289, [%rd67];
|
|
ld.local.f32 %f290, [%rd67+4];
|
|
mul.f32 %f291, %f290, 0f3F372474;
|
|
fma.rn.f32 %f292, %f289, 0f3E59999A, %f291;
|
|
ld.local.f32 %f293, [%rd67+8];
|
|
fma.rn.f32 %f294, %f293, 0f3D93A92A, %f292;
|
|
fma.rn.f32 %f684, %f276, %f294, %f684;
|
|
fma.rn.f32 %f683, %f277, %f294, %f683;
|
|
fma.rn.f32 %f682, %f278, %f294, %f682;
|
|
mul.f32 %f295, %f671, %f277;
|
|
fma.rn.f32 %f296, %f670, %f276, %f295;
|
|
fma.rn.f32 %f297, %f672, %f278, %f296;
|
|
cvt.sat.f32.f32 %f298, %f297;
|
|
fma.rn.f32 %f681, %f298, %f289, %f681;
|
|
fma.rn.f32 %f680, %f298, %f290, %f680;
|
|
fma.rn.f32 %f679, %f298, %f293, %f679;
|
|
add.s32 %r304, %r304, 1;
|
|
setp.lt.s32 %p38, %r304, %r1;
|
|
@%p38 bra BB0_6;
|
|
|
|
add.s32 %r302, %r302, 1;
|
|
setp.lt.s32 %p39, %r302, %r1;
|
|
@%p39 bra BB0_5;
|
|
|
|
BB0_54:
|
|
cvt.rn.f32.u32 %f299, %r4;
|
|
cvt.rn.f32.u32 %f300, %r3;
|
|
tex.2d.v4.f32.f32 {%f301, %f302, %f303, %f304}, [albedoTex, {%f300, %f299}];
|
|
mul.lo.s32 %r231, %r1, %r1;
|
|
cvt.rn.f32.s32 %f305, %r231;
|
|
rcp.rn.f32 %f306, %f305;
|
|
mul.f32 %f307, %f681, %f306;
|
|
mul.f32 %f308, %f680, %f306;
|
|
mul.f32 %f309, %f679, %f306;
|
|
fma.rn.f32 %f310, %f681, %f306, %f307;
|
|
fma.rn.f32 %f311, %f680, %f306, %f308;
|
|
fma.rn.f32 %f312, %f679, %f306, %f309;
|
|
mul.f32 %f87, %f310, %f301;
|
|
mul.f32 %f88, %f311, %f302;
|
|
mul.f32 %f89, %f312, %f303;
|
|
add.f32 %f90, %f159, %f87;
|
|
add.f32 %f91, %f160, %f88;
|
|
add.f32 %f92, %f161, %f89;
|
|
ld.global.u32 %r327, [imageEnabled];
|
|
and.b32 %r232, %r327, 1;
|
|
setp.eq.b32 %p40, %r232, 1;
|
|
@!%p40 bra BB0_89;
|
|
bra.uni BB0_55;
|
|
|
|
BB0_55:
|
|
abs.f32 %f94, %f90;
|
|
setp.lt.f32 %p41, %f94, 0f00800000;
|
|
mul.f32 %f318, %f94, 0f4B800000;
|
|
selp.f32 %f319, 0fC3170000, 0fC2FE0000, %p41;
|
|
selp.f32 %f320, %f318, %f94, %p41;
|
|
mov.b32 %r233, %f320;
|
|
and.b32 %r234, %r233, 8388607;
|
|
or.b32 %r235, %r234, 1065353216;
|
|
mov.b32 %f321, %r235;
|
|
shr.u32 %r236, %r233, 23;
|
|
cvt.rn.f32.u32 %f322, %r236;
|
|
add.f32 %f323, %f319, %f322;
|
|
setp.gt.f32 %p42, %f321, 0f3FB504F3;
|
|
mul.f32 %f324, %f321, 0f3F000000;
|
|
add.f32 %f325, %f323, 0f3F800000;
|
|
selp.f32 %f326, %f324, %f321, %p42;
|
|
selp.f32 %f327, %f325, %f323, %p42;
|
|
add.f32 %f328, %f326, 0fBF800000;
|
|
add.f32 %f314, %f326, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f313,%f314;
|
|
// inline asm
|
|
add.f32 %f329, %f328, %f328;
|
|
mul.f32 %f330, %f313, %f329;
|
|
mul.f32 %f331, %f330, %f330;
|
|
mov.f32 %f332, 0f3C4CAF63;
|
|
mov.f32 %f333, 0f3B18F0FE;
|
|
fma.rn.f32 %f334, %f333, %f331, %f332;
|
|
mov.f32 %f335, 0f3DAAAABD;
|
|
fma.rn.f32 %f336, %f334, %f331, %f335;
|
|
mul.rn.f32 %f337, %f336, %f331;
|
|
mul.rn.f32 %f338, %f337, %f330;
|
|
sub.f32 %f339, %f328, %f330;
|
|
neg.f32 %f340, %f330;
|
|
add.f32 %f341, %f339, %f339;
|
|
fma.rn.f32 %f342, %f340, %f328, %f341;
|
|
mul.rn.f32 %f343, %f313, %f342;
|
|
add.f32 %f344, %f338, %f330;
|
|
sub.f32 %f345, %f330, %f344;
|
|
add.f32 %f346, %f338, %f345;
|
|
add.f32 %f347, %f343, %f346;
|
|
add.f32 %f348, %f344, %f347;
|
|
sub.f32 %f349, %f344, %f348;
|
|
add.f32 %f350, %f347, %f349;
|
|
mov.f32 %f351, 0f3F317200;
|
|
mul.rn.f32 %f352, %f327, %f351;
|
|
mov.f32 %f353, 0f35BFBE8E;
|
|
mul.rn.f32 %f354, %f327, %f353;
|
|
add.f32 %f355, %f352, %f348;
|
|
sub.f32 %f356, %f352, %f355;
|
|
add.f32 %f357, %f348, %f356;
|
|
add.f32 %f358, %f350, %f357;
|
|
add.f32 %f359, %f354, %f358;
|
|
add.f32 %f360, %f355, %f359;
|
|
sub.f32 %f361, %f355, %f360;
|
|
add.f32 %f362, %f359, %f361;
|
|
mov.f32 %f363, 0f3EE66666;
|
|
mul.rn.f32 %f364, %f363, %f360;
|
|
neg.f32 %f365, %f364;
|
|
fma.rn.f32 %f366, %f363, %f360, %f365;
|
|
fma.rn.f32 %f367, %f363, %f362, %f366;
|
|
mov.f32 %f368, 0f00000000;
|
|
fma.rn.f32 %f369, %f368, %f360, %f367;
|
|
add.rn.f32 %f370, %f364, %f369;
|
|
neg.f32 %f371, %f370;
|
|
add.rn.f32 %f372, %f364, %f371;
|
|
add.rn.f32 %f373, %f372, %f369;
|
|
mov.b32 %r237, %f370;
|
|
setp.eq.s32 %p43, %r237, 1118925336;
|
|
add.s32 %r238, %r237, -1;
|
|
mov.b32 %f374, %r238;
|
|
add.f32 %f375, %f373, 0f37000000;
|
|
selp.f32 %f376, %f374, %f370, %p43;
|
|
selp.f32 %f95, %f375, %f373, %p43;
|
|
mul.f32 %f377, %f376, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f378, %f377;
|
|
mov.f32 %f379, 0fBF317200;
|
|
fma.rn.f32 %f380, %f378, %f379, %f376;
|
|
mov.f32 %f381, 0fB5BFBE8E;
|
|
fma.rn.f32 %f382, %f378, %f381, %f380;
|
|
mul.f32 %f383, %f382, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f384, %f383;
|
|
add.f32 %f385, %f378, 0f00000000;
|
|
ex2.approx.f32 %f386, %f385;
|
|
mul.f32 %f387, %f384, %f386;
|
|
setp.lt.f32 %p44, %f376, 0fC2D20000;
|
|
selp.f32 %f388, 0f00000000, %f387, %p44;
|
|
setp.gt.f32 %p45, %f376, 0f42D20000;
|
|
selp.f32 %f703, 0f7F800000, %f388, %p45;
|
|
setp.eq.f32 %p46, %f703, 0f7F800000;
|
|
@%p46 bra BB0_57;
|
|
|
|
fma.rn.f32 %f703, %f703, %f95, %f703;
|
|
|
|
BB0_57:
|
|
mov.f32 %f624, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f623, %f624;
|
|
fma.rn.f32 %f622, %f623, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f621, %f622;
|
|
setp.lt.f32 %p47, %f90, 0f00000000;
|
|
setp.eq.f32 %p48, %f621, 0f3F800000;
|
|
and.pred %p1, %p47, %p48;
|
|
mov.b32 %r239, %f703;
|
|
xor.b32 %r240, %r239, -2147483648;
|
|
mov.b32 %f389, %r240;
|
|
selp.f32 %f705, %f389, %f703, %p1;
|
|
setp.eq.f32 %p49, %f90, 0f00000000;
|
|
@%p49 bra BB0_60;
|
|
bra.uni BB0_58;
|
|
|
|
BB0_60:
|
|
add.f32 %f392, %f90, %f90;
|
|
selp.f32 %f705, %f392, 0f00000000, %p48;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_58:
|
|
setp.geu.f32 %p50, %f90, 0f00000000;
|
|
@%p50 bra BB0_61;
|
|
|
|
mov.f32 %f651, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f391, %f651;
|
|
setp.neu.f32 %p51, %f391, 0f3EE66666;
|
|
selp.f32 %f705, 0f7FFFFFFF, %f705, %p51;
|
|
|
|
BB0_61:
|
|
add.f32 %f626, %f159, %f87;
|
|
abs.f32 %f625, %f626;
|
|
add.f32 %f393, %f625, 0f3EE66666;
|
|
mov.b32 %r241, %f393;
|
|
setp.lt.s32 %p53, %r241, 2139095040;
|
|
@%p53 bra BB0_66;
|
|
|
|
add.f32 %f648, %f159, %f87;
|
|
abs.f32 %f647, %f648;
|
|
setp.gtu.f32 %p54, %f647, 0f7F800000;
|
|
@%p54 bra BB0_65;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_65:
|
|
add.f32 %f705, %f90, 0f3EE66666;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_63:
|
|
add.f32 %f650, %f159, %f87;
|
|
abs.f32 %f649, %f650;
|
|
setp.neu.f32 %p55, %f649, 0f7F800000;
|
|
@%p55 bra BB0_66;
|
|
|
|
selp.f32 %f705, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_66:
|
|
mov.f32 %f635, 0fB5BFBE8E;
|
|
mov.f32 %f634, 0fBF317200;
|
|
mov.f32 %f633, 0f00000000;
|
|
mov.f32 %f632, 0f35BFBE8E;
|
|
mov.f32 %f631, 0f3F317200;
|
|
mov.f32 %f630, 0f3DAAAABD;
|
|
mov.f32 %f629, 0f3C4CAF63;
|
|
mov.f32 %f628, 0f3B18F0FE;
|
|
mov.f32 %f627, 0f3EE66666;
|
|
setp.eq.f32 %p56, %f90, 0f3F800000;
|
|
selp.f32 %f106, 0f3F800000, %f705, %p56;
|
|
abs.f32 %f107, %f91;
|
|
setp.lt.f32 %p57, %f107, 0f00800000;
|
|
mul.f32 %f396, %f107, 0f4B800000;
|
|
selp.f32 %f397, 0fC3170000, 0fC2FE0000, %p57;
|
|
selp.f32 %f398, %f396, %f107, %p57;
|
|
mov.b32 %r242, %f398;
|
|
and.b32 %r243, %r242, 8388607;
|
|
or.b32 %r244, %r243, 1065353216;
|
|
mov.b32 %f399, %r244;
|
|
shr.u32 %r245, %r242, 23;
|
|
cvt.rn.f32.u32 %f400, %r245;
|
|
add.f32 %f401, %f397, %f400;
|
|
setp.gt.f32 %p58, %f399, 0f3FB504F3;
|
|
mul.f32 %f402, %f399, 0f3F000000;
|
|
add.f32 %f403, %f401, 0f3F800000;
|
|
selp.f32 %f404, %f402, %f399, %p58;
|
|
selp.f32 %f405, %f403, %f401, %p58;
|
|
add.f32 %f406, %f404, 0fBF800000;
|
|
add.f32 %f395, %f404, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f394,%f395;
|
|
// inline asm
|
|
add.f32 %f407, %f406, %f406;
|
|
mul.f32 %f408, %f394, %f407;
|
|
mul.f32 %f409, %f408, %f408;
|
|
fma.rn.f32 %f412, %f628, %f409, %f629;
|
|
fma.rn.f32 %f414, %f412, %f409, %f630;
|
|
mul.rn.f32 %f415, %f414, %f409;
|
|
mul.rn.f32 %f416, %f415, %f408;
|
|
sub.f32 %f417, %f406, %f408;
|
|
neg.f32 %f418, %f408;
|
|
add.f32 %f419, %f417, %f417;
|
|
fma.rn.f32 %f420, %f418, %f406, %f419;
|
|
mul.rn.f32 %f421, %f394, %f420;
|
|
add.f32 %f422, %f416, %f408;
|
|
sub.f32 %f423, %f408, %f422;
|
|
add.f32 %f424, %f416, %f423;
|
|
add.f32 %f425, %f421, %f424;
|
|
add.f32 %f426, %f422, %f425;
|
|
sub.f32 %f427, %f422, %f426;
|
|
add.f32 %f428, %f425, %f427;
|
|
mul.rn.f32 %f430, %f405, %f631;
|
|
mul.rn.f32 %f432, %f405, %f632;
|
|
add.f32 %f433, %f430, %f426;
|
|
sub.f32 %f434, %f430, %f433;
|
|
add.f32 %f435, %f426, %f434;
|
|
add.f32 %f436, %f428, %f435;
|
|
add.f32 %f437, %f432, %f436;
|
|
add.f32 %f438, %f433, %f437;
|
|
sub.f32 %f439, %f433, %f438;
|
|
add.f32 %f440, %f437, %f439;
|
|
mul.rn.f32 %f442, %f627, %f438;
|
|
neg.f32 %f443, %f442;
|
|
fma.rn.f32 %f444, %f627, %f438, %f443;
|
|
fma.rn.f32 %f445, %f627, %f440, %f444;
|
|
fma.rn.f32 %f447, %f633, %f438, %f445;
|
|
add.rn.f32 %f448, %f442, %f447;
|
|
neg.f32 %f449, %f448;
|
|
add.rn.f32 %f450, %f442, %f449;
|
|
add.rn.f32 %f451, %f450, %f447;
|
|
mov.b32 %r246, %f448;
|
|
setp.eq.s32 %p59, %r246, 1118925336;
|
|
add.s32 %r247, %r246, -1;
|
|
mov.b32 %f452, %r247;
|
|
add.f32 %f453, %f451, 0f37000000;
|
|
selp.f32 %f454, %f452, %f448, %p59;
|
|
selp.f32 %f108, %f453, %f451, %p59;
|
|
mul.f32 %f455, %f454, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f456, %f455;
|
|
fma.rn.f32 %f458, %f456, %f634, %f454;
|
|
fma.rn.f32 %f460, %f456, %f635, %f458;
|
|
mul.f32 %f461, %f460, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f462, %f461;
|
|
add.f32 %f463, %f456, 0f00000000;
|
|
ex2.approx.f32 %f464, %f463;
|
|
mul.f32 %f465, %f462, %f464;
|
|
setp.lt.f32 %p60, %f454, 0fC2D20000;
|
|
selp.f32 %f466, 0f00000000, %f465, %p60;
|
|
setp.gt.f32 %p61, %f454, 0f42D20000;
|
|
selp.f32 %f706, 0f7F800000, %f466, %p61;
|
|
setp.eq.f32 %p62, %f706, 0f7F800000;
|
|
@%p62 bra BB0_68;
|
|
|
|
fma.rn.f32 %f706, %f706, %f108, %f706;
|
|
|
|
BB0_68:
|
|
setp.lt.f32 %p63, %f91, 0f00000000;
|
|
and.pred %p2, %p63, %p48;
|
|
mov.b32 %r248, %f706;
|
|
xor.b32 %r249, %r248, -2147483648;
|
|
mov.b32 %f467, %r249;
|
|
selp.f32 %f708, %f467, %f706, %p2;
|
|
setp.eq.f32 %p65, %f91, 0f00000000;
|
|
@%p65 bra BB0_71;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_71:
|
|
add.f32 %f470, %f91, %f91;
|
|
selp.f32 %f708, %f470, 0f00000000, %p48;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_69:
|
|
setp.geu.f32 %p66, %f91, 0f00000000;
|
|
@%p66 bra BB0_72;
|
|
|
|
mov.f32 %f646, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f469, %f646;
|
|
setp.neu.f32 %p67, %f469, 0f3EE66666;
|
|
selp.f32 %f708, 0f7FFFFFFF, %f708, %p67;
|
|
|
|
BB0_72:
|
|
add.f32 %f653, %f160, %f88;
|
|
abs.f32 %f652, %f653;
|
|
add.f32 %f471, %f652, 0f3EE66666;
|
|
mov.b32 %r250, %f471;
|
|
setp.lt.s32 %p69, %r250, 2139095040;
|
|
@%p69 bra BB0_77;
|
|
|
|
add.f32 %f656, %f160, %f88;
|
|
abs.f32 %f655, %f656;
|
|
setp.gtu.f32 %p70, %f655, 0f7F800000;
|
|
@%p70 bra BB0_76;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_76:
|
|
add.f32 %f708, %f91, 0f3EE66666;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_74:
|
|
add.f32 %f658, %f160, %f88;
|
|
abs.f32 %f657, %f658;
|
|
setp.neu.f32 %p71, %f657, 0f7F800000;
|
|
@%p71 bra BB0_77;
|
|
|
|
selp.f32 %f708, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_77:
|
|
mov.f32 %f644, 0fB5BFBE8E;
|
|
mov.f32 %f643, 0fBF317200;
|
|
mov.f32 %f642, 0f00000000;
|
|
mov.f32 %f641, 0f35BFBE8E;
|
|
mov.f32 %f640, 0f3F317200;
|
|
mov.f32 %f639, 0f3DAAAABD;
|
|
mov.f32 %f638, 0f3C4CAF63;
|
|
mov.f32 %f637, 0f3B18F0FE;
|
|
mov.f32 %f636, 0f3EE66666;
|
|
setp.eq.f32 %p72, %f91, 0f3F800000;
|
|
selp.f32 %f119, 0f3F800000, %f708, %p72;
|
|
abs.f32 %f120, %f92;
|
|
setp.lt.f32 %p73, %f120, 0f00800000;
|
|
mul.f32 %f474, %f120, 0f4B800000;
|
|
selp.f32 %f475, 0fC3170000, 0fC2FE0000, %p73;
|
|
selp.f32 %f476, %f474, %f120, %p73;
|
|
mov.b32 %r251, %f476;
|
|
and.b32 %r252, %r251, 8388607;
|
|
or.b32 %r253, %r252, 1065353216;
|
|
mov.b32 %f477, %r253;
|
|
shr.u32 %r254, %r251, 23;
|
|
cvt.rn.f32.u32 %f478, %r254;
|
|
add.f32 %f479, %f475, %f478;
|
|
setp.gt.f32 %p74, %f477, 0f3FB504F3;
|
|
mul.f32 %f480, %f477, 0f3F000000;
|
|
add.f32 %f481, %f479, 0f3F800000;
|
|
selp.f32 %f482, %f480, %f477, %p74;
|
|
selp.f32 %f483, %f481, %f479, %p74;
|
|
add.f32 %f484, %f482, 0fBF800000;
|
|
add.f32 %f473, %f482, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f472,%f473;
|
|
// inline asm
|
|
add.f32 %f485, %f484, %f484;
|
|
mul.f32 %f486, %f472, %f485;
|
|
mul.f32 %f487, %f486, %f486;
|
|
fma.rn.f32 %f490, %f637, %f487, %f638;
|
|
fma.rn.f32 %f492, %f490, %f487, %f639;
|
|
mul.rn.f32 %f493, %f492, %f487;
|
|
mul.rn.f32 %f494, %f493, %f486;
|
|
sub.f32 %f495, %f484, %f486;
|
|
neg.f32 %f496, %f486;
|
|
add.f32 %f497, %f495, %f495;
|
|
fma.rn.f32 %f498, %f496, %f484, %f497;
|
|
mul.rn.f32 %f499, %f472, %f498;
|
|
add.f32 %f500, %f494, %f486;
|
|
sub.f32 %f501, %f486, %f500;
|
|
add.f32 %f502, %f494, %f501;
|
|
add.f32 %f503, %f499, %f502;
|
|
add.f32 %f504, %f500, %f503;
|
|
sub.f32 %f505, %f500, %f504;
|
|
add.f32 %f506, %f503, %f505;
|
|
mul.rn.f32 %f508, %f483, %f640;
|
|
mul.rn.f32 %f510, %f483, %f641;
|
|
add.f32 %f511, %f508, %f504;
|
|
sub.f32 %f512, %f508, %f511;
|
|
add.f32 %f513, %f504, %f512;
|
|
add.f32 %f514, %f506, %f513;
|
|
add.f32 %f515, %f510, %f514;
|
|
add.f32 %f516, %f511, %f515;
|
|
sub.f32 %f517, %f511, %f516;
|
|
add.f32 %f518, %f515, %f517;
|
|
mul.rn.f32 %f520, %f636, %f516;
|
|
neg.f32 %f521, %f520;
|
|
fma.rn.f32 %f522, %f636, %f516, %f521;
|
|
fma.rn.f32 %f523, %f636, %f518, %f522;
|
|
fma.rn.f32 %f525, %f642, %f516, %f523;
|
|
add.rn.f32 %f526, %f520, %f525;
|
|
neg.f32 %f527, %f526;
|
|
add.rn.f32 %f528, %f520, %f527;
|
|
add.rn.f32 %f529, %f528, %f525;
|
|
mov.b32 %r255, %f526;
|
|
setp.eq.s32 %p75, %r255, 1118925336;
|
|
add.s32 %r256, %r255, -1;
|
|
mov.b32 %f530, %r256;
|
|
add.f32 %f531, %f529, 0f37000000;
|
|
selp.f32 %f532, %f530, %f526, %p75;
|
|
selp.f32 %f121, %f531, %f529, %p75;
|
|
mul.f32 %f533, %f532, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f534, %f533;
|
|
fma.rn.f32 %f536, %f534, %f643, %f532;
|
|
fma.rn.f32 %f538, %f534, %f644, %f536;
|
|
mul.f32 %f539, %f538, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f540, %f539;
|
|
add.f32 %f541, %f534, 0f00000000;
|
|
ex2.approx.f32 %f542, %f541;
|
|
mul.f32 %f543, %f540, %f542;
|
|
setp.lt.f32 %p76, %f532, 0fC2D20000;
|
|
selp.f32 %f544, 0f00000000, %f543, %p76;
|
|
setp.gt.f32 %p77, %f532, 0f42D20000;
|
|
selp.f32 %f709, 0f7F800000, %f544, %p77;
|
|
setp.eq.f32 %p78, %f709, 0f7F800000;
|
|
@%p78 bra BB0_79;
|
|
|
|
fma.rn.f32 %f709, %f709, %f121, %f709;
|
|
|
|
BB0_79:
|
|
setp.lt.f32 %p79, %f92, 0f00000000;
|
|
and.pred %p3, %p79, %p48;
|
|
mov.b32 %r257, %f709;
|
|
xor.b32 %r258, %r257, -2147483648;
|
|
mov.b32 %f545, %r258;
|
|
selp.f32 %f711, %f545, %f709, %p3;
|
|
setp.eq.f32 %p81, %f92, 0f00000000;
|
|
@%p81 bra BB0_82;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_82:
|
|
add.f32 %f548, %f92, %f92;
|
|
selp.f32 %f711, %f548, 0f00000000, %p48;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_80:
|
|
setp.geu.f32 %p82, %f92, 0f00000000;
|
|
@%p82 bra BB0_83;
|
|
|
|
mov.f32 %f645, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f547, %f645;
|
|
setp.neu.f32 %p83, %f547, 0f3EE66666;
|
|
selp.f32 %f711, 0f7FFFFFFF, %f711, %p83;
|
|
|
|
BB0_83:
|
|
add.f32 %f660, %f161, %f89;
|
|
abs.f32 %f659, %f660;
|
|
add.f32 %f549, %f659, 0f3EE66666;
|
|
mov.b32 %r259, %f549;
|
|
setp.lt.s32 %p85, %r259, 2139095040;
|
|
@%p85 bra BB0_88;
|
|
|
|
add.f32 %f663, %f161, %f89;
|
|
abs.f32 %f662, %f663;
|
|
setp.gtu.f32 %p86, %f662, 0f7F800000;
|
|
@%p86 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f711, %f92, 0f3EE66666;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
add.f32 %f665, %f161, %f89;
|
|
abs.f32 %f664, %f665;
|
|
setp.neu.f32 %p87, %f664, 0f7F800000;
|
|
@%p87 bra BB0_88;
|
|
|
|
selp.f32 %f711, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_88:
|
|
mov.u32 %r293, 4;
|
|
mov.u64 %rd132, 0;
|
|
mov.u32 %r292, 2;
|
|
setp.eq.f32 %p88, %f92, 0f3F800000;
|
|
selp.f32 %f550, 0f3F800000, %f711, %p88;
|
|
cvt.u64.u32 %rd72, %r4;
|
|
cvt.u64.u32 %rd71, %r3;
|
|
mov.u64 %rd75, image;
|
|
cvta.global.u64 %rd70, %rd75;
|
|
// inline asm
|
|
call (%rd69), _rt_buffer_get_64, (%rd70, %r292, %r293, %rd71, %rd72, %rd132, %rd132);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f551, %f550;
|
|
mul.f32 %f552, %f551, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r262, %f552;
|
|
cvt.sat.f32.f32 %f553, %f119;
|
|
mul.f32 %f554, %f553, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r263, %f554;
|
|
cvt.sat.f32.f32 %f555, %f106;
|
|
mul.f32 %f556, %f555, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r264, %f556;
|
|
cvt.u16.u32 %rs16, %r262;
|
|
cvt.u16.u32 %rs17, %r264;
|
|
cvt.u16.u32 %rs18, %r263;
|
|
mov.u16 %rs19, 255;
|
|
st.v4.u8 [%rd69], {%rs16, %rs18, %rs17, %rs19};
|
|
ld.global.u32 %r327, [imageEnabled];
|
|
|
|
BB0_89:
|
|
and.b32 %r265, %r327, 4;
|
|
setp.eq.s32 %p89, %r265, 0;
|
|
@%p89 bra BB0_91;
|
|
|
|
add.f32 %f661, %f160, %f88;
|
|
add.f32 %f654, %f159, %f87;
|
|
mov.u32 %r295, 8;
|
|
mov.u64 %rd133, 0;
|
|
mov.u32 %r294, 2;
|
|
cvt.u64.u32 %rd78, %r3;
|
|
cvt.u64.u32 %rd79, %r4;
|
|
mov.u64 %rd82, image_HDR;
|
|
cvta.global.u64 %rd77, %rd82;
|
|
// inline asm
|
|
call (%rd76), _rt_buffer_get_64, (%rd77, %r294, %r295, %rd78, %rd79, %rd133, %rd133);
|
|
// inline asm
|
|
mov.f32 %f560, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f560;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f92;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f661;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f654;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd76], {%rs20, %rs21, %rs22, %rs23};
|
|
ld.global.u32 %r327, [imageEnabled];
|
|
|
|
BB0_91:
|
|
and.b32 %r268, %r327, 16;
|
|
setp.eq.s32 %p90, %r268, 0;
|
|
@%p90 bra BB0_93;
|
|
|
|
mov.u32 %r297, 8;
|
|
mov.u64 %rd134, 0;
|
|
mov.u32 %r296, 2;
|
|
cvt.u64.u32 %rd85, %r3;
|
|
cvt.u64.u32 %rd86, %r4;
|
|
mov.u64 %rd89, image_HDR2;
|
|
cvta.global.u64 %rd84, %rd89;
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd84, %r296, %r297, %rd85, %rd86, %rd134, %rd134);
|
|
// inline asm
|
|
mov.f32 %f564, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f564;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs24, %f87;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs24, %rs25, %rs26, %rs27};
|
|
ld.global.u32 %r327, [imageEnabled];
|
|
|
|
BB0_93:
|
|
and.b32 %r271, %r327, 64;
|
|
setp.eq.s32 %p91, %r271, 0;
|
|
@%p91 bra BB0_109;
|
|
|
|
mov.u32 %r299, 4;
|
|
mov.u64 %rd135, 0;
|
|
mov.u32 %r298, 2;
|
|
mul.f32 %f565, %f160, 0f3F372474;
|
|
fma.rn.f32 %f566, %f159, 0f3E59999A, %f565;
|
|
fma.rn.f32 %f132, %f161, 0f3D93A92A, %f566;
|
|
cvt.u64.u32 %rd93, %r4;
|
|
cvt.u64.u32 %rd92, %r3;
|
|
mov.u64 %rd96, lightmapDirectDir;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r298, %r299, %rd92, %rd93, %rd135, %rd135);
|
|
// inline asm
|
|
ld.v4.u8 {%rs28, %rs29, %rs30, %rs31}, [%rd90];
|
|
cvt.rn.f32.u16 %f567, %rs28;
|
|
div.rn.f32 %f568, %f567, 0f437F0000;
|
|
fma.rn.f32 %f569, %f568, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f570, %rs29;
|
|
div.rn.f32 %f571, %f570, 0f437F0000;
|
|
fma.rn.f32 %f572, %f571, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f573, %rs30;
|
|
div.rn.f32 %f574, %f573, 0f437F0000;
|
|
fma.rn.f32 %f575, %f574, 0f40000000, 0fBF800000;
|
|
mul.f32 %f576, %f572, %f572;
|
|
fma.rn.f32 %f577, %f569, %f569, %f576;
|
|
fma.rn.f32 %f578, %f575, %f575, %f577;
|
|
sqrt.rn.f32 %f579, %f578;
|
|
rcp.rn.f32 %f580, %f579;
|
|
mul.f32 %f712, %f569, %f580;
|
|
mul.f32 %f713, %f572, %f580;
|
|
mul.f32 %f714, %f575, %f580;
|
|
abs.f32 %f581, %f684;
|
|
setp.gt.f32 %p92, %f581, 0f00000000;
|
|
@%p92 bra BB0_97;
|
|
|
|
abs.f32 %f582, %f683;
|
|
setp.gt.f32 %p93, %f582, 0f00000000;
|
|
@%p93 bra BB0_97;
|
|
|
|
abs.f32 %f583, %f682;
|
|
setp.leu.f32 %p94, %f583, 0f00000000;
|
|
@%p94 bra BB0_100;
|
|
|
|
BB0_97:
|
|
ld.global.u8 %rs32, [imageEnabled];
|
|
and.b16 %rs33, %rs32, 16;
|
|
setp.ne.s16 %p95, %rs33, 0;
|
|
@%p95 bra BB0_98;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_98:
|
|
mov.f32 %f712, %f684;
|
|
mov.f32 %f713, %f683;
|
|
mov.f32 %f714, %f682;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_99:
|
|
mul.f32 %f584, %f683, %f683;
|
|
fma.rn.f32 %f585, %f684, %f684, %f584;
|
|
fma.rn.f32 %f586, %f682, %f682, %f585;
|
|
sqrt.rn.f32 %f587, %f586;
|
|
rcp.rn.f32 %f588, %f587;
|
|
mul.f32 %f589, %f684, %f588;
|
|
mul.f32 %f590, %f683, %f588;
|
|
mul.f32 %f591, %f682, %f588;
|
|
mul.f32 %f592, %f88, 0f3F372474;
|
|
fma.rn.f32 %f593, %f87, 0f3E59999A, %f592;
|
|
fma.rn.f32 %f594, %f89, 0f3D93A92A, %f593;
|
|
mul.f32 %f595, %f594, %f589;
|
|
mul.f32 %f596, %f594, %f590;
|
|
mul.f32 %f597, %f594, %f591;
|
|
fma.rn.f32 %f712, %f132, %f712, %f595;
|
|
fma.rn.f32 %f713, %f132, %f713, %f596;
|
|
fma.rn.f32 %f714, %f132, %f714, %f597;
|
|
|
|
BB0_100:
|
|
mov.u32 %r301, 4;
|
|
mov.u64 %rd136, 0;
|
|
mov.u32 %r300, 2;
|
|
mul.f32 %f598, %f713, %f713;
|
|
fma.rn.f32 %f599, %f712, %f712, %f598;
|
|
fma.rn.f32 %f600, %f714, %f714, %f599;
|
|
sqrt.rn.f32 %f601, %f600;
|
|
rcp.rn.f32 %f602, %f601;
|
|
mul.f32 %f603, %f712, %f602;
|
|
mul.f32 %f604, %f713, %f602;
|
|
mul.f32 %f605, %f714, %f602;
|
|
mul.f32 %f606, %f671, %f604;
|
|
fma.rn.f32 %f607, %f670, %f603, %f606;
|
|
fma.rn.f32 %f608, %f672, %f605, %f607;
|
|
fma.rn.f32 %f609, %f608, 0f3F000000, 0f3F000000;
|
|
mov.f32 %f610, 0f3B808081;
|
|
max.f32 %f611, %f609, %f610;
|
|
mul.f32 %f612, %f611, 0f437F0000;
|
|
cvt.rzi.s32.f32 %r276, %f612;
|
|
mov.u64 %rd103, image_Dir;
|
|
cvta.global.u64 %rd98, %rd103;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r300, %r301, %rd92, %rd93, %rd136, %rd136);
|
|
// inline asm
|
|
fma.rn.f32 %f613, %f603, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f614, %f613, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r277, %f614;
|
|
fma.rn.f32 %f615, %f604, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f616, %f615, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r278, %f616;
|
|
fma.rn.f32 %f617, %f605, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f618, %f617, 0f437F0000;
|
|
cvt.rzi.u32.f32 %r279, %f618;
|
|
cvt.u16.u32 %rs34, %r276;
|
|
cvt.u16.u32 %rs35, %r279;
|
|
cvt.u16.u32 %rs36, %r278;
|
|
cvt.u16.u32 %rs37, %r277;
|
|
st.v4.u8 [%rd97], {%rs37, %rs36, %rs35, %rs34};
|
|
|
|
BB0_109:
|
|
ret;
|
|
}
|
|
|
|
|