1914 lines
52 KiB
Plaintext
1914 lines
52 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_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .texref sky;
|
|
.global .align 4 .b8 skyColor[12];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .u32 hemispherical;
|
|
.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_typeinfo8skyColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo13hemisphericalE[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_typename8skyColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename13hemisphericalE[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_typeenum8skyColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13hemisphericalE = 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_semantic8skyColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13hemisphericalE[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_annotation8skyColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13hemisphericalE[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[32];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<105>;
|
|
.reg .b16 %rs<148>;
|
|
.reg .f32 %f<708>;
|
|
.reg .b32 %r<356>;
|
|
.reg .b64 %rd<283>;
|
|
|
|
|
|
mov.u64 %rd282, __local_depot0;
|
|
cvta.local.u64 %SP, %rd282;
|
|
ld.global.u32 %r1, [samples];
|
|
shl.b32 %r2, %r1, 1;
|
|
ld.global.v2.u32 {%r94, %r95}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r94;
|
|
cvt.u64.u32 %rd23, %r95;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r92, 2;
|
|
mov.u32 %r93, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r92, %r93, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r3, [%rd20];
|
|
shr.u32 %r98, %r3, 16;
|
|
cvt.u16.u32 %rs1, %r98;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r3;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p4, %rs9, 0;
|
|
mov.f32 %f664, 0f00000000;
|
|
mov.f32 %f665, %f664;
|
|
mov.f32 %f666, %f664;
|
|
@%p4 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd20+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f143, %rs12;
|
|
div.rn.f32 %f144, %f143, 0f437F0000;
|
|
fma.rn.f32 %f145, %f144, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f146, %rs10;
|
|
div.rn.f32 %f147, %f146, 0f437F0000;
|
|
fma.rn.f32 %f148, %f147, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f149, %rs7;
|
|
div.rn.f32 %f150, %f149, 0f437F0000;
|
|
fma.rn.f32 %f151, %f150, 0f40000000, 0fBF800000;
|
|
mul.f32 %f152, %f148, %f148;
|
|
fma.rn.f32 %f153, %f145, %f145, %f152;
|
|
fma.rn.f32 %f154, %f151, %f151, %f153;
|
|
sqrt.rn.f32 %f155, %f154;
|
|
rcp.rn.f32 %f156, %f155;
|
|
mul.f32 %f664, %f145, %f156;
|
|
mul.f32 %f665, %f148, %f156;
|
|
mul.f32 %f666, %f151, %f156;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r99, %r100}, [pixelID];
|
|
ld.global.v2.u32 {%r102, %r103}, [tileInfo];
|
|
add.s32 %r4, %r99, %r102;
|
|
add.s32 %r5, %r100, %r103;
|
|
setp.eq.f32 %p5, %f665, 0f00000000;
|
|
setp.eq.f32 %p6, %f664, 0f00000000;
|
|
and.pred %p7, %p6, %p5;
|
|
setp.eq.f32 %p8, %f666, 0f00000000;
|
|
and.pred %p9, %p7, %p8;
|
|
@%p9 bra BB0_107;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_107:
|
|
ld.global.u32 %r355, [imageEnabled];
|
|
and.b32 %r290, %r355, 1;
|
|
setp.eq.b32 %p98, %r290, 1;
|
|
@!%p98 bra BB0_109;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_108:
|
|
cvt.u64.u32 %rd173, %r5;
|
|
cvt.u64.u32 %rd172, %r4;
|
|
mov.u64 %rd176, image;
|
|
cvta.global.u64 %rd171, %rd176;
|
|
mov.u64 %rd175, 0;
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd171, %r92, %r93, %rd172, %rd173, %rd175, %rd175);
|
|
// inline asm
|
|
mov.u16 %rs82, 0;
|
|
st.v4.u8 [%rd170], {%rs82, %rs82, %rs82, %rs82};
|
|
ld.global.u32 %r355, [imageEnabled];
|
|
|
|
BB0_109:
|
|
cvt.u64.u32 %rd18, %r4;
|
|
cvt.u64.u32 %rd19, %r5;
|
|
and.b32 %r293, %r355, 4;
|
|
setp.eq.s32 %p99, %r293, 0;
|
|
@%p99 bra BB0_113;
|
|
|
|
ld.global.u32 %r294, [additive];
|
|
setp.eq.s32 %p100, %r294, 0;
|
|
@%p100 bra BB0_112;
|
|
|
|
mov.u64 %rd189, image_HDR;
|
|
cvta.global.u64 %rd178, %rd189;
|
|
mov.u32 %r298, 8;
|
|
mov.u64 %rd188, 0;
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r92, %r298, %rd18, %rd19, %rd188, %rd188);
|
|
// inline asm
|
|
ld.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd177];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f593, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f594, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f595, %rs91;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd183), _rt_buffer_get_64, (%rd178, %r92, %r298, %rd18, %rd19, %rd188, %rd188);
|
|
// inline asm
|
|
add.f32 %f596, %f593, 0f00000000;
|
|
add.f32 %f597, %f594, 0f00000000;
|
|
add.f32 %f598, %f595, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f598;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f597;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f596;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs93, 0;
|
|
st.v4.u16 [%rd183], {%rs86, %rs87, %rs88, %rs93};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r111, %r112}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r111;
|
|
cvt.u64.u32 %rd30, %r112;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r108, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r92, %r108, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f162, [%rd27+8];
|
|
ld.f32 %f163, [%rd27+4];
|
|
ld.f32 %f164, [%rd27];
|
|
mul.f32 %f165, %f164, 0f3456BF95;
|
|
mul.f32 %f166, %f163, 0f3456BF95;
|
|
mul.f32 %f167, %f162, 0f3456BF95;
|
|
abs.f32 %f168, %f664;
|
|
div.rn.f32 %f169, %f165, %f168;
|
|
abs.f32 %f170, %f665;
|
|
div.rn.f32 %f171, %f166, %f170;
|
|
abs.f32 %f172, %f666;
|
|
div.rn.f32 %f173, %f167, %f172;
|
|
abs.f32 %f174, %f169;
|
|
abs.f32 %f175, %f171;
|
|
abs.f32 %f176, %f173;
|
|
mov.f32 %f177, 0f38D1B717;
|
|
max.f32 %f178, %f174, %f177;
|
|
max.f32 %f179, %f175, %f177;
|
|
max.f32 %f180, %f176, %f177;
|
|
fma.rn.f32 %f7, %f664, %f178, %f164;
|
|
fma.rn.f32 %f8, %f665, %f179, %f163;
|
|
fma.rn.f32 %f9, %f666, %f180, %f162;
|
|
ld.global.u32 %r6, [hemispherical];
|
|
setp.gt.f32 %p10, %f168, %f172;
|
|
neg.f32 %f181, %f665;
|
|
selp.f32 %f182, %f181, 0f00000000, %p10;
|
|
neg.f32 %f183, %f666;
|
|
selp.f32 %f184, %f664, %f183, %p10;
|
|
selp.f32 %f185, 0f00000000, %f665, %p10;
|
|
mul.f32 %f186, %f184, %f184;
|
|
fma.rn.f32 %f187, %f182, %f182, %f186;
|
|
fma.rn.f32 %f188, %f185, %f185, %f187;
|
|
sqrt.rn.f32 %f189, %f188;
|
|
rcp.rn.f32 %f190, %f189;
|
|
mul.f32 %f10, %f182, %f190;
|
|
mul.f32 %f11, %f184, %f190;
|
|
mul.f32 %f12, %f185, %f190;
|
|
ld.global.v2.u32 {%r115, %r116}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r115;
|
|
cvt.u64.u32 %rd36, %r116;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r92, %r93, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f693, 0f00000000;
|
|
setp.lt.s32 %p11, %r1, 1;
|
|
mov.f32 %f692, %f693;
|
|
mov.f32 %f691, %f693;
|
|
mov.f32 %f690, %f693;
|
|
mov.f32 %f689, %f693;
|
|
@%p11 bra BB0_56;
|
|
|
|
cvt.rn.f32.s32 %f196, %r2;
|
|
rcp.rn.f32 %f13, %f196;
|
|
ld.u32 %r333, [%rd33];
|
|
mul.f32 %f14, %f7, 0f3456BF95;
|
|
mul.f32 %f15, %f8, 0f3456BF95;
|
|
mul.f32 %f16, %f9, 0f3456BF95;
|
|
mul.f32 %f197, %f664, %f11;
|
|
mul.f32 %f198, %f665, %f10;
|
|
sub.f32 %f17, %f198, %f197;
|
|
mul.f32 %f199, %f666, %f10;
|
|
mul.f32 %f200, %f664, %f12;
|
|
sub.f32 %f18, %f200, %f199;
|
|
mul.f32 %f201, %f665, %f12;
|
|
mul.f32 %f202, %f666, %f11;
|
|
sub.f32 %f19, %f202, %f201;
|
|
mov.f32 %f693, 0f00000000;
|
|
mov.u32 %r119, 0;
|
|
abs.f32 %f277, %f15;
|
|
abs.f32 %f278, %f14;
|
|
max.f32 %f279, %f278, %f277;
|
|
abs.f32 %f280, %f16;
|
|
max.f32 %f281, %f279, %f280;
|
|
mov.u32 %r330, %r119;
|
|
mov.f32 %f692, %f693;
|
|
mov.f32 %f691, %f693;
|
|
mov.f32 %f690, %f693;
|
|
mov.f32 %f689, %f693;
|
|
|
|
BB0_5:
|
|
mov.u32 %r332, %r119;
|
|
|
|
BB0_6:
|
|
cvt.rn.f32.s32 %f662, %r330;
|
|
mad.lo.s32 %r121, %r333, 1664525, 1013904223;
|
|
and.b32 %r122, %r121, 16777215;
|
|
cvt.rn.f32.u32 %f203, %r122;
|
|
fma.rn.f32 %f204, %f203, 0f33800000, %f662;
|
|
mul.f32 %f205, %f13, %f204;
|
|
mad.lo.s32 %r333, %r121, 1664525, 1013904223;
|
|
and.b32 %r123, %r333, 16777215;
|
|
cvt.rn.f32.u32 %f206, %r123;
|
|
cvt.rn.f32.s32 %f207, %r332;
|
|
fma.rn.f32 %f208, %f206, 0f33800000, %f207;
|
|
mul.f32 %f209, %f13, %f208;
|
|
fma.rn.f32 %f31, %f205, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f210, %f31, %f31;
|
|
mov.f32 %f211, 0f3F800000;
|
|
sub.f32 %f212, %f211, %f210;
|
|
mov.f32 %f213, 0f00000000;
|
|
max.f32 %f214, %f213, %f212;
|
|
sqrt.rn.f32 %f32, %f214;
|
|
mul.f32 %f683, %f209, 0f40C90FDB;
|
|
abs.f32 %f34, %f683;
|
|
setp.neu.f32 %p12, %f34, 0f7F800000;
|
|
mov.f32 %f677, %f683;
|
|
@%p12 bra BB0_8;
|
|
|
|
mul.rn.f32 %f677, %f683, %f213;
|
|
|
|
BB0_8:
|
|
mul.f32 %f216, %f677, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r343, %f216;
|
|
cvt.rn.f32.s32 %f217, %r343;
|
|
neg.f32 %f218, %f217;
|
|
mov.f32 %f219, 0f3FC90FDA;
|
|
fma.rn.f32 %f220, %f218, %f219, %f677;
|
|
mov.f32 %f221, 0f33A22168;
|
|
fma.rn.f32 %f222, %f218, %f221, %f220;
|
|
mov.f32 %f223, 0f27C234C5;
|
|
fma.rn.f32 %f678, %f218, %f223, %f222;
|
|
abs.f32 %f224, %f677;
|
|
setp.leu.f32 %p13, %f224, 0f47CE4780;
|
|
@%p13 bra BB0_19;
|
|
|
|
add.u64 %rd42, %SP, 4;
|
|
cvta.to.local.u64 %rd3, %rd42;
|
|
mov.b32 %r14, %f677;
|
|
shr.u32 %r15, %r14, 23;
|
|
shl.b32 %r126, %r14, 8;
|
|
or.b32 %r16, %r126, -2147483648;
|
|
mov.u32 %r334, 0;
|
|
mov.u64 %rd279, 0;
|
|
mov.u64 %rd278, %rd3;
|
|
mov.u32 %r335, %r334;
|
|
|
|
BB0_10:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd43, %rd279, 2;
|
|
mov.u64 %rd44, __cudart_i2opi_f;
|
|
add.s64 %rd45, %rd44, %rd43;
|
|
ld.const.u32 %r129, [%rd45];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r127, %r129, %r16, %r335;
|
|
madc.hi.u32 %r335, %r129, %r16, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd278], %r127;
|
|
add.s32 %r334, %r334, 1;
|
|
cvt.s64.s32 %rd279, %r334;
|
|
mul.wide.s32 %rd48, %r334, 4;
|
|
add.s64 %rd278, %rd3, %rd48;
|
|
setp.ne.s32 %p14, %r334, 6;
|
|
@%p14 bra BB0_10;
|
|
|
|
add.u64 %rd277, %SP, 4;
|
|
and.b32 %r132, %r15, 255;
|
|
add.s32 %r133, %r132, -128;
|
|
shr.u32 %r134, %r133, 5;
|
|
and.b32 %r21, %r14, -2147483648;
|
|
cvta.to.local.u64 %rd50, %rd277;
|
|
st.local.u32 [%rd50+24], %r335;
|
|
mov.u32 %r135, 6;
|
|
sub.s32 %r136, %r135, %r134;
|
|
mul.wide.s32 %rd51, %r136, 4;
|
|
add.s64 %rd8, %rd50, %rd51;
|
|
ld.local.u32 %r336, [%rd8];
|
|
ld.local.u32 %r337, [%rd8+-4];
|
|
and.b32 %r24, %r15, 31;
|
|
setp.eq.s32 %p15, %r24, 0;
|
|
@%p15 bra BB0_13;
|
|
|
|
mov.u32 %r137, 32;
|
|
sub.s32 %r138, %r137, %r24;
|
|
shr.u32 %r139, %r337, %r138;
|
|
shl.b32 %r140, %r336, %r24;
|
|
add.s32 %r336, %r139, %r140;
|
|
ld.local.u32 %r141, [%rd8+-8];
|
|
shr.u32 %r142, %r141, %r138;
|
|
shl.b32 %r143, %r337, %r24;
|
|
add.s32 %r337, %r142, %r143;
|
|
|
|
BB0_13:
|
|
shr.u32 %r144, %r337, 30;
|
|
shl.b32 %r145, %r336, 2;
|
|
add.s32 %r338, %r144, %r145;
|
|
shl.b32 %r30, %r337, 2;
|
|
shr.u32 %r146, %r338, 31;
|
|
shr.u32 %r147, %r336, 30;
|
|
add.s32 %r31, %r146, %r147;
|
|
setp.eq.s32 %p16, %r146, 0;
|
|
@%p16 bra BB0_14;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_14:
|
|
mov.u32 %r339, %r21;
|
|
mov.u32 %r340, %r30;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
not.b32 %r148, %r338;
|
|
neg.s32 %r340, %r30;
|
|
setp.eq.s32 %p17, %r30, 0;
|
|
selp.u32 %r149, 1, 0, %p17;
|
|
add.s32 %r338, %r149, %r148;
|
|
xor.b32 %r339, %r21, -2147483648;
|
|
|
|
BB0_16:
|
|
clz.b32 %r342, %r338;
|
|
setp.eq.s32 %p18, %r342, 0;
|
|
shl.b32 %r150, %r338, %r342;
|
|
mov.u32 %r151, 32;
|
|
sub.s32 %r152, %r151, %r342;
|
|
shr.u32 %r153, %r340, %r152;
|
|
add.s32 %r154, %r153, %r150;
|
|
selp.b32 %r39, %r338, %r154, %p18;
|
|
mov.u32 %r155, -921707870;
|
|
mul.hi.u32 %r341, %r39, %r155;
|
|
setp.eq.s32 %p19, %r21, 0;
|
|
neg.s32 %r156, %r31;
|
|
selp.b32 %r343, %r31, %r156, %p19;
|
|
setp.lt.s32 %p20, %r341, 1;
|
|
@%p20 bra BB0_18;
|
|
|
|
mul.lo.s32 %r157, %r39, -921707870;
|
|
shr.u32 %r158, %r157, 31;
|
|
shl.b32 %r159, %r341, 1;
|
|
add.s32 %r341, %r158, %r159;
|
|
add.s32 %r342, %r342, 1;
|
|
|
|
BB0_18:
|
|
mov.u32 %r160, 126;
|
|
sub.s32 %r161, %r160, %r342;
|
|
shl.b32 %r162, %r161, 23;
|
|
add.s32 %r163, %r341, 1;
|
|
shr.u32 %r164, %r163, 7;
|
|
add.s32 %r165, %r164, 1;
|
|
shr.u32 %r166, %r165, 1;
|
|
add.s32 %r167, %r166, %r162;
|
|
or.b32 %r168, %r167, %r339;
|
|
mov.b32 %f678, %r168;
|
|
|
|
BB0_19:
|
|
mul.rn.f32 %f40, %f678, %f678;
|
|
add.s32 %r47, %r343, 1;
|
|
and.b32 %r48, %r47, 1;
|
|
setp.eq.s32 %p21, %r48, 0;
|
|
@%p21 bra BB0_21;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_21:
|
|
mov.f32 %f227, 0f3C08839E;
|
|
mov.f32 %f228, 0fB94CA1F9;
|
|
fma.rn.f32 %f679, %f228, %f40, %f227;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_20:
|
|
mov.f32 %f225, 0fBAB6061A;
|
|
mov.f32 %f226, 0f37CCF5CE;
|
|
fma.rn.f32 %f679, %f226, %f40, %f225;
|
|
|
|
BB0_22:
|
|
@%p21 bra BB0_24;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_24:
|
|
mov.f32 %f232, 0fBE2AAAA3;
|
|
fma.rn.f32 %f233, %f679, %f40, %f232;
|
|
fma.rn.f32 %f680, %f233, %f40, %f213;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_23:
|
|
mov.f32 %f229, 0f3D2AAAA5;
|
|
fma.rn.f32 %f230, %f679, %f40, %f229;
|
|
mov.f32 %f231, 0fBF000000;
|
|
fma.rn.f32 %f680, %f230, %f40, %f231;
|
|
|
|
BB0_25:
|
|
fma.rn.f32 %f681, %f680, %f678, %f678;
|
|
@%p21 bra BB0_27;
|
|
|
|
fma.rn.f32 %f681, %f680, %f40, %f211;
|
|
|
|
BB0_27:
|
|
and.b32 %r169, %r47, 2;
|
|
setp.eq.s32 %p24, %r169, 0;
|
|
@%p24 bra BB0_29;
|
|
|
|
mov.f32 %f237, 0fBF800000;
|
|
fma.rn.f32 %f681, %f681, %f237, %f213;
|
|
|
|
BB0_29:
|
|
@%p12 bra BB0_31;
|
|
|
|
mul.rn.f32 %f683, %f683, %f213;
|
|
|
|
BB0_31:
|
|
mov.f32 %f663, 0f3FC90FDA;
|
|
mul.f32 %f239, %f683, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r353, %f239;
|
|
cvt.rn.f32.s32 %f240, %r353;
|
|
neg.f32 %f241, %f240;
|
|
fma.rn.f32 %f243, %f241, %f663, %f683;
|
|
fma.rn.f32 %f245, %f241, %f221, %f243;
|
|
fma.rn.f32 %f684, %f241, %f223, %f245;
|
|
abs.f32 %f247, %f683;
|
|
setp.leu.f32 %p26, %f247, 0f47CE4780;
|
|
@%p26 bra BB0_42;
|
|
|
|
add.u64 %rd53, %SP, 4;
|
|
cvta.to.local.u64 %rd9, %rd53;
|
|
mov.b32 %r50, %f683;
|
|
shr.u32 %r51, %r50, 23;
|
|
shl.b32 %r172, %r50, 8;
|
|
or.b32 %r52, %r172, -2147483648;
|
|
mov.u32 %r344, 0;
|
|
mov.u64 %rd280, %rd9;
|
|
mov.u64 %rd281, %rd25;
|
|
mov.u32 %r345, %r344;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd54, %rd281, 2;
|
|
mov.u64 %rd55, __cudart_i2opi_f;
|
|
add.s64 %rd56, %rd55, %rd54;
|
|
ld.const.u32 %r175, [%rd56];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r173, %r175, %r52, %r345;
|
|
madc.hi.u32 %r345, %r175, %r52, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd280], %r173;
|
|
add.s32 %r344, %r344, 1;
|
|
cvt.s64.s32 %rd281, %r344;
|
|
mul.wide.s32 %rd57, %r344, 4;
|
|
add.s64 %rd280, %rd9, %rd57;
|
|
setp.ne.s32 %p27, %r344, 6;
|
|
@%p27 bra BB0_33;
|
|
|
|
and.b32 %r178, %r51, 255;
|
|
add.s32 %r179, %r178, -128;
|
|
shr.u32 %r180, %r179, 5;
|
|
and.b32 %r57, %r50, -2147483648;
|
|
cvta.to.local.u64 %rd59, %rd53;
|
|
st.local.u32 [%rd59+24], %r345;
|
|
mov.u32 %r181, 6;
|
|
sub.s32 %r182, %r181, %r180;
|
|
mul.wide.s32 %rd60, %r182, 4;
|
|
add.s64 %rd15, %rd59, %rd60;
|
|
ld.local.u32 %r346, [%rd15];
|
|
ld.local.u32 %r347, [%rd15+-4];
|
|
and.b32 %r60, %r51, 31;
|
|
setp.eq.s32 %p28, %r60, 0;
|
|
@%p28 bra BB0_36;
|
|
|
|
mov.u32 %r183, 32;
|
|
sub.s32 %r184, %r183, %r60;
|
|
shr.u32 %r185, %r347, %r184;
|
|
shl.b32 %r186, %r346, %r60;
|
|
add.s32 %r346, %r185, %r186;
|
|
ld.local.u32 %r187, [%rd15+-8];
|
|
shr.u32 %r188, %r187, %r184;
|
|
shl.b32 %r189, %r347, %r60;
|
|
add.s32 %r347, %r188, %r189;
|
|
|
|
BB0_36:
|
|
shr.u32 %r190, %r347, 30;
|
|
shl.b32 %r191, %r346, 2;
|
|
add.s32 %r348, %r190, %r191;
|
|
shl.b32 %r66, %r347, 2;
|
|
shr.u32 %r192, %r348, 31;
|
|
shr.u32 %r193, %r346, 30;
|
|
add.s32 %r67, %r192, %r193;
|
|
setp.eq.s32 %p29, %r192, 0;
|
|
@%p29 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r349, %r57;
|
|
mov.u32 %r350, %r66;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r194, %r348;
|
|
neg.s32 %r350, %r66;
|
|
setp.eq.s32 %p30, %r66, 0;
|
|
selp.u32 %r195, 1, 0, %p30;
|
|
add.s32 %r348, %r195, %r194;
|
|
xor.b32 %r349, %r57, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r352, %r348;
|
|
setp.eq.s32 %p31, %r352, 0;
|
|
shl.b32 %r196, %r348, %r352;
|
|
mov.u32 %r197, 32;
|
|
sub.s32 %r198, %r197, %r352;
|
|
shr.u32 %r199, %r350, %r198;
|
|
add.s32 %r200, %r199, %r196;
|
|
selp.b32 %r75, %r348, %r200, %p31;
|
|
mov.u32 %r201, -921707870;
|
|
mul.hi.u32 %r351, %r75, %r201;
|
|
setp.eq.s32 %p32, %r57, 0;
|
|
neg.s32 %r202, %r67;
|
|
selp.b32 %r353, %r67, %r202, %p32;
|
|
setp.lt.s32 %p33, %r351, 1;
|
|
@%p33 bra BB0_41;
|
|
|
|
mul.lo.s32 %r203, %r75, -921707870;
|
|
shr.u32 %r204, %r203, 31;
|
|
shl.b32 %r205, %r351, 1;
|
|
add.s32 %r351, %r204, %r205;
|
|
add.s32 %r352, %r352, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r206, 126;
|
|
sub.s32 %r207, %r206, %r352;
|
|
shl.b32 %r208, %r207, 23;
|
|
add.s32 %r209, %r351, 1;
|
|
shr.u32 %r210, %r209, 7;
|
|
add.s32 %r211, %r210, 1;
|
|
shr.u32 %r212, %r211, 1;
|
|
add.s32 %r213, %r212, %r208;
|
|
or.b32 %r214, %r213, %r349;
|
|
mov.b32 %f684, %r214;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f57, %f684, %f684;
|
|
and.b32 %r83, %r353, 1;
|
|
setp.eq.s32 %p34, %r83, 0;
|
|
@%p34 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f250, 0f3C08839E;
|
|
mov.f32 %f251, 0fB94CA1F9;
|
|
fma.rn.f32 %f685, %f251, %f57, %f250;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f248, 0fBAB6061A;
|
|
mov.f32 %f249, 0f37CCF5CE;
|
|
fma.rn.f32 %f685, %f249, %f57, %f248;
|
|
|
|
BB0_45:
|
|
@%p34 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f255, 0fBE2AAAA3;
|
|
fma.rn.f32 %f256, %f685, %f57, %f255;
|
|
fma.rn.f32 %f686, %f256, %f57, %f213;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f252, 0f3D2AAAA5;
|
|
fma.rn.f32 %f253, %f685, %f57, %f252;
|
|
mov.f32 %f254, 0fBF000000;
|
|
fma.rn.f32 %f686, %f253, %f57, %f254;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f687, %f686, %f684, %f684;
|
|
@%p34 bra BB0_50;
|
|
|
|
fma.rn.f32 %f687, %f686, %f57, %f211;
|
|
|
|
BB0_50:
|
|
and.b32 %r215, %r353, 2;
|
|
setp.eq.s32 %p37, %r215, 0;
|
|
@%p37 bra BB0_52;
|
|
|
|
mov.f32 %f260, 0fBF800000;
|
|
fma.rn.f32 %f687, %f687, %f260, %f213;
|
|
|
|
BB0_52:
|
|
mul.f32 %f261, %f32, %f681;
|
|
mul.f32 %f262, %f32, %f687;
|
|
mul.f32 %f263, %f10, %f262;
|
|
mul.f32 %f264, %f11, %f262;
|
|
mul.f32 %f265, %f12, %f262;
|
|
fma.rn.f32 %f266, %f19, %f261, %f263;
|
|
fma.rn.f32 %f267, %f18, %f261, %f264;
|
|
fma.rn.f32 %f268, %f17, %f261, %f265;
|
|
fma.rn.f32 %f69, %f664, %f31, %f266;
|
|
fma.rn.f32 %f70, %f665, %f31, %f267;
|
|
fma.rn.f32 %f71, %f666, %f31, %f268;
|
|
setp.gt.f32 %p38, %f70, 0f00000000;
|
|
setp.eq.s32 %p39, %r6, 0;
|
|
or.pred %p40, %p39, %p38;
|
|
@!%p40 bra BB0_54;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_53:
|
|
add.u64 %rd61, %SP, 0;
|
|
cvta.to.local.u64 %rd62, %rd61;
|
|
max.f32 %f275, %f281, %f177;
|
|
mov.u32 %r219, 1065353216;
|
|
st.local.u32 [%rd62], %r219;
|
|
ld.global.u32 %r216, [root];
|
|
mov.u32 %r217, 1;
|
|
mov.f32 %f276, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r216, %f7, %f8, %f9, %f69, %f70, %f71, %r217, %f275, %f276, %rd61, %r93);
|
|
// inline asm
|
|
mul.f32 %f283, %f665, %f70;
|
|
fma.rn.f32 %f284, %f664, %f69, %f283;
|
|
fma.rn.f32 %f285, %f666, %f71, %f284;
|
|
ld.local.f32 %f286, [%rd62];
|
|
fma.rn.f32 %f691, %f69, %f286, %f691;
|
|
fma.rn.f32 %f692, %f70, %f286, %f692;
|
|
fma.rn.f32 %f693, %f71, %f286, %f693;
|
|
add.f32 %f690, %f690, %f286;
|
|
cvt.sat.f32.f32 %f287, %f285;
|
|
fma.rn.f32 %f689, %f287, %f286, %f689;
|
|
|
|
BB0_54:
|
|
add.s32 %r332, %r332, 1;
|
|
setp.lt.s32 %p41, %r332, %r2;
|
|
@%p41 bra BB0_6;
|
|
|
|
add.s32 %r330, %r330, 1;
|
|
setp.lt.s32 %p42, %r330, %r2;
|
|
@%p42 bra BB0_5;
|
|
|
|
BB0_56:
|
|
mul.lo.s32 %r220, %r2, %r2;
|
|
cvt.rn.f32.s32 %f288, %r220;
|
|
div.rn.f32 %f289, %f689, %f288;
|
|
div.rn.f32 %f87, %f690, %f288;
|
|
div.rn.f32 %f88, %f691, %f288;
|
|
div.rn.f32 %f89, %f692, %f288;
|
|
div.rn.f32 %f90, %f693, %f288;
|
|
add.f32 %f290, %f289, %f289;
|
|
ld.global.f32 %f291, [skyColor];
|
|
mul.f32 %f91, %f290, %f291;
|
|
ld.global.f32 %f292, [skyColor+4];
|
|
mul.f32 %f92, %f290, %f292;
|
|
ld.global.f32 %f293, [skyColor+8];
|
|
mul.f32 %f93, %f290, %f293;
|
|
ld.global.u32 %r354, [imageEnabled];
|
|
and.b32 %r221, %r354, 1;
|
|
setp.eq.b32 %p43, %r221, 1;
|
|
@!%p43 bra BB0_91;
|
|
bra.uni BB0_57;
|
|
|
|
BB0_57:
|
|
abs.f32 %f95, %f91;
|
|
setp.lt.f32 %p44, %f95, 0f00800000;
|
|
mul.f32 %f299, %f95, 0f4B800000;
|
|
selp.f32 %f300, 0fC3170000, 0fC2FE0000, %p44;
|
|
selp.f32 %f301, %f299, %f95, %p44;
|
|
mov.b32 %r222, %f301;
|
|
and.b32 %r223, %r222, 8388607;
|
|
or.b32 %r224, %r223, 1065353216;
|
|
mov.b32 %f302, %r224;
|
|
shr.u32 %r225, %r222, 23;
|
|
cvt.rn.f32.u32 %f303, %r225;
|
|
add.f32 %f304, %f300, %f303;
|
|
setp.gt.f32 %p45, %f302, 0f3FB504F3;
|
|
mul.f32 %f305, %f302, 0f3F000000;
|
|
add.f32 %f306, %f304, 0f3F800000;
|
|
selp.f32 %f307, %f305, %f302, %p45;
|
|
selp.f32 %f308, %f306, %f304, %p45;
|
|
add.f32 %f309, %f307, 0fBF800000;
|
|
add.f32 %f295, %f307, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f294,%f295;
|
|
// inline asm
|
|
add.f32 %f310, %f309, %f309;
|
|
mul.f32 %f311, %f294, %f310;
|
|
mul.f32 %f312, %f311, %f311;
|
|
mov.f32 %f313, 0f3C4CAF63;
|
|
mov.f32 %f314, 0f3B18F0FE;
|
|
fma.rn.f32 %f315, %f314, %f312, %f313;
|
|
mov.f32 %f316, 0f3DAAAABD;
|
|
fma.rn.f32 %f317, %f315, %f312, %f316;
|
|
mul.rn.f32 %f318, %f317, %f312;
|
|
mul.rn.f32 %f319, %f318, %f311;
|
|
sub.f32 %f320, %f309, %f311;
|
|
neg.f32 %f321, %f311;
|
|
add.f32 %f322, %f320, %f320;
|
|
fma.rn.f32 %f323, %f321, %f309, %f322;
|
|
mul.rn.f32 %f324, %f294, %f323;
|
|
add.f32 %f325, %f319, %f311;
|
|
sub.f32 %f326, %f311, %f325;
|
|
add.f32 %f327, %f319, %f326;
|
|
add.f32 %f328, %f324, %f327;
|
|
add.f32 %f329, %f325, %f328;
|
|
sub.f32 %f330, %f325, %f329;
|
|
add.f32 %f331, %f328, %f330;
|
|
mov.f32 %f332, 0f3F317200;
|
|
mul.rn.f32 %f333, %f308, %f332;
|
|
mov.f32 %f334, 0f35BFBE8E;
|
|
mul.rn.f32 %f335, %f308, %f334;
|
|
add.f32 %f336, %f333, %f329;
|
|
sub.f32 %f337, %f333, %f336;
|
|
add.f32 %f338, %f329, %f337;
|
|
add.f32 %f339, %f331, %f338;
|
|
add.f32 %f340, %f335, %f339;
|
|
add.f32 %f341, %f336, %f340;
|
|
sub.f32 %f342, %f336, %f341;
|
|
add.f32 %f343, %f340, %f342;
|
|
mov.f32 %f344, 0f3EE66666;
|
|
mul.rn.f32 %f345, %f344, %f341;
|
|
neg.f32 %f346, %f345;
|
|
fma.rn.f32 %f347, %f344, %f341, %f346;
|
|
fma.rn.f32 %f348, %f344, %f343, %f347;
|
|
mov.f32 %f349, 0f00000000;
|
|
fma.rn.f32 %f350, %f349, %f341, %f348;
|
|
add.rn.f32 %f351, %f345, %f350;
|
|
neg.f32 %f352, %f351;
|
|
add.rn.f32 %f353, %f345, %f352;
|
|
add.rn.f32 %f354, %f353, %f350;
|
|
mov.b32 %r226, %f351;
|
|
setp.eq.s32 %p46, %r226, 1118925336;
|
|
add.s32 %r227, %r226, -1;
|
|
mov.b32 %f355, %r227;
|
|
add.f32 %f356, %f354, 0f37000000;
|
|
selp.f32 %f357, %f355, %f351, %p46;
|
|
selp.f32 %f96, %f356, %f354, %p46;
|
|
mul.f32 %f358, %f357, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f359, %f358;
|
|
mov.f32 %f360, 0fBF317200;
|
|
fma.rn.f32 %f361, %f359, %f360, %f357;
|
|
mov.f32 %f362, 0fB5BFBE8E;
|
|
fma.rn.f32 %f363, %f359, %f362, %f361;
|
|
mul.f32 %f364, %f363, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f365, %f364;
|
|
add.f32 %f366, %f359, 0f00000000;
|
|
ex2.approx.f32 %f367, %f366;
|
|
mul.f32 %f368, %f365, %f367;
|
|
setp.lt.f32 %p47, %f357, 0fC2D20000;
|
|
selp.f32 %f369, 0f00000000, %f368, %p47;
|
|
setp.gt.f32 %p48, %f357, 0f42D20000;
|
|
selp.f32 %f699, 0f7F800000, %f369, %p48;
|
|
setp.eq.f32 %p49, %f699, 0f7F800000;
|
|
@%p49 bra BB0_59;
|
|
|
|
fma.rn.f32 %f699, %f699, %f96, %f699;
|
|
|
|
BB0_59:
|
|
mov.f32 %f631, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f630, %f631;
|
|
fma.rn.f32 %f629, %f630, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f628, %f629;
|
|
setp.lt.f32 %p50, %f91, 0f00000000;
|
|
setp.eq.f32 %p51, %f628, 0f3F800000;
|
|
and.pred %p1, %p50, %p51;
|
|
mov.b32 %r228, %f699;
|
|
xor.b32 %r229, %r228, -2147483648;
|
|
mov.b32 %f370, %r229;
|
|
selp.f32 %f701, %f370, %f699, %p1;
|
|
setp.eq.f32 %p52, %f91, 0f00000000;
|
|
@%p52 bra BB0_62;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_62:
|
|
add.f32 %f373, %f91, %f91;
|
|
selp.f32 %f701, %f373, 0f00000000, %p51;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_112:
|
|
mov.u64 %rd196, image_HDR;
|
|
cvta.global.u64 %rd191, %rd196;
|
|
mov.u32 %r300, 8;
|
|
mov.u64 %rd195, 0;
|
|
// inline asm
|
|
call (%rd190), _rt_buffer_get_64, (%rd191, %r92, %r300, %rd18, %rd19, %rd195, %rd195);
|
|
// inline asm
|
|
mov.f32 %f599, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f599;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs95, 0;
|
|
st.v4.u16 [%rd190], {%rs94, %rs94, %rs94, %rs95};
|
|
|
|
BB0_113:
|
|
ld.global.u32 %r301, [additive];
|
|
setp.eq.s32 %p101, %r301, 0;
|
|
@%p101 bra BB0_115;
|
|
|
|
mov.u64 %rd209, image_RNM0;
|
|
cvta.global.u64 %rd198, %rd209;
|
|
mov.u32 %r305, 8;
|
|
mov.u64 %rd208, 0;
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd198, %r92, %r305, %rd18, %rd19, %rd208, %rd208);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd197];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f600, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f601, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f602, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd203), _rt_buffer_get_64, (%rd198, %r92, %r305, %rd18, %rd19, %rd208, %rd208);
|
|
// inline asm
|
|
add.f32 %f603, %f600, 0f00000000;
|
|
add.f32 %f604, %f601, 0f00000000;
|
|
add.f32 %f605, %f602, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f605;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f604;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f603;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd203], {%rs99, %rs100, %rs101, %rs106};
|
|
bra.uni BB0_116;
|
|
|
|
BB0_115:
|
|
mov.u64 %rd216, image_RNM0;
|
|
cvta.global.u64 %rd211, %rd216;
|
|
mov.u32 %r307, 8;
|
|
mov.u64 %rd215, 0;
|
|
// inline asm
|
|
call (%rd210), _rt_buffer_get_64, (%rd211, %r92, %r307, %rd18, %rd19, %rd215, %rd215);
|
|
// inline asm
|
|
mov.f32 %f606, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f606;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs108, 0;
|
|
st.v4.u16 [%rd210], {%rs107, %rs107, %rs107, %rs108};
|
|
|
|
BB0_116:
|
|
ld.global.u32 %r308, [additive];
|
|
setp.eq.s32 %p102, %r308, 0;
|
|
@%p102 bra BB0_118;
|
|
|
|
mov.u64 %rd229, image_RNM1;
|
|
cvta.global.u64 %rd218, %rd229;
|
|
mov.u32 %r312, 8;
|
|
mov.u64 %rd228, 0;
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd218, %r92, %r312, %rd18, %rd19, %rd228, %rd228);
|
|
// inline asm
|
|
ld.v4.u16 {%rs115, %rs116, %rs117, %rs118}, [%rd217];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f607, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f608, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f609, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd223), _rt_buffer_get_64, (%rd218, %r92, %r312, %rd18, %rd19, %rd228, %rd228);
|
|
// inline asm
|
|
add.f32 %f610, %f607, 0f00000000;
|
|
add.f32 %f611, %f608, 0f00000000;
|
|
add.f32 %f612, %f609, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f612;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f611;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f610;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd223], {%rs112, %rs113, %rs114, %rs119};
|
|
bra.uni BB0_119;
|
|
|
|
BB0_118:
|
|
mov.u64 %rd236, image_RNM1;
|
|
cvta.global.u64 %rd231, %rd236;
|
|
mov.u32 %r314, 8;
|
|
mov.u64 %rd235, 0;
|
|
// inline asm
|
|
call (%rd230), _rt_buffer_get_64, (%rd231, %r92, %r314, %rd18, %rd19, %rd235, %rd235);
|
|
// inline asm
|
|
mov.f32 %f613, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f613;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs121, 0;
|
|
st.v4.u16 [%rd230], {%rs120, %rs120, %rs120, %rs121};
|
|
|
|
BB0_119:
|
|
ld.global.u32 %r315, [additive];
|
|
setp.eq.s32 %p103, %r315, 0;
|
|
@%p103 bra BB0_121;
|
|
|
|
mov.u64 %rd249, image_RNM2;
|
|
cvta.global.u64 %rd238, %rd249;
|
|
mov.u32 %r319, 8;
|
|
mov.u64 %rd248, 0;
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd238, %r92, %r319, %rd18, %rd19, %rd248, %rd248);
|
|
// inline asm
|
|
ld.v4.u16 {%rs128, %rs129, %rs130, %rs131}, [%rd237];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f614, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f615, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f616, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd243), _rt_buffer_get_64, (%rd238, %r92, %r319, %rd18, %rd19, %rd248, %rd248);
|
|
// inline asm
|
|
add.f32 %f617, %f614, 0f00000000;
|
|
add.f32 %f618, %f615, 0f00000000;
|
|
add.f32 %f619, %f616, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f619;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f618;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f617;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd243], {%rs125, %rs126, %rs127, %rs132};
|
|
bra.uni BB0_122;
|
|
|
|
BB0_121:
|
|
mov.u64 %rd256, image_RNM2;
|
|
cvta.global.u64 %rd251, %rd256;
|
|
mov.u32 %r321, 8;
|
|
mov.u64 %rd255, 0;
|
|
// inline asm
|
|
call (%rd250), _rt_buffer_get_64, (%rd251, %r92, %r321, %rd18, %rd19, %rd255, %rd255);
|
|
// inline asm
|
|
mov.f32 %f620, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f620;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs134, 0;
|
|
st.v4.u16 [%rd250], {%rs133, %rs133, %rs133, %rs134};
|
|
|
|
BB0_122:
|
|
ld.global.u32 %r322, [additive];
|
|
setp.eq.s32 %p104, %r322, 0;
|
|
@%p104 bra BB0_124;
|
|
|
|
mov.u64 %rd269, image_RNM3;
|
|
cvta.global.u64 %rd258, %rd269;
|
|
mov.u32 %r326, 8;
|
|
mov.u64 %rd268, 0;
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd258, %r92, %r326, %rd18, %rd19, %rd268, %rd268);
|
|
// inline asm
|
|
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd257];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f621, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f622, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f623, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd263), _rt_buffer_get_64, (%rd258, %r92, %r326, %rd18, %rd19, %rd268, %rd268);
|
|
// inline asm
|
|
add.f32 %f624, %f621, 0f00000000;
|
|
add.f32 %f625, %f622, 0f00000000;
|
|
add.f32 %f626, %f623, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f626;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f625;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f624;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd263], {%rs138, %rs139, %rs140, %rs145};
|
|
bra.uni BB0_125;
|
|
|
|
BB0_124:
|
|
mov.u64 %rd276, image_RNM3;
|
|
cvta.global.u64 %rd271, %rd276;
|
|
mov.u32 %r328, 8;
|
|
mov.u64 %rd275, 0;
|
|
// inline asm
|
|
call (%rd270), _rt_buffer_get_64, (%rd271, %r92, %r328, %rd18, %rd19, %rd275, %rd275);
|
|
// inline asm
|
|
mov.f32 %f627, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f627;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs147, 0;
|
|
st.v4.u16 [%rd270], {%rs146, %rs146, %rs146, %rs147};
|
|
bra.uni BB0_125;
|
|
|
|
BB0_60:
|
|
setp.geu.f32 %p53, %f91, 0f00000000;
|
|
@%p53 bra BB0_63;
|
|
|
|
mov.f32 %f655, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f372, %f655;
|
|
setp.neu.f32 %p54, %f372, 0f3EE66666;
|
|
selp.f32 %f701, 0f7FFFFFFF, %f701, %p54;
|
|
|
|
BB0_63:
|
|
abs.f32 %f632, %f91;
|
|
add.f32 %f374, %f632, 0f3EE66666;
|
|
mov.b32 %r230, %f374;
|
|
setp.lt.s32 %p56, %r230, 2139095040;
|
|
@%p56 bra BB0_68;
|
|
|
|
abs.f32 %f653, %f91;
|
|
setp.gtu.f32 %p57, %f653, 0f7F800000;
|
|
@%p57 bra BB0_67;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_67:
|
|
add.f32 %f701, %f91, 0f3EE66666;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_65:
|
|
abs.f32 %f654, %f91;
|
|
setp.neu.f32 %p58, %f654, 0f7F800000;
|
|
@%p58 bra BB0_68;
|
|
|
|
selp.f32 %f701, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_68:
|
|
mov.f32 %f641, 0fB5BFBE8E;
|
|
mov.f32 %f640, 0fBF317200;
|
|
mov.f32 %f639, 0f00000000;
|
|
mov.f32 %f638, 0f35BFBE8E;
|
|
mov.f32 %f637, 0f3F317200;
|
|
mov.f32 %f636, 0f3DAAAABD;
|
|
mov.f32 %f635, 0f3C4CAF63;
|
|
mov.f32 %f634, 0f3B18F0FE;
|
|
mov.f32 %f633, 0f3EE66666;
|
|
setp.eq.f32 %p59, %f91, 0f3F800000;
|
|
selp.f32 %f107, 0f3F800000, %f701, %p59;
|
|
abs.f32 %f108, %f92;
|
|
setp.lt.f32 %p60, %f108, 0f00800000;
|
|
mul.f32 %f377, %f108, 0f4B800000;
|
|
selp.f32 %f378, 0fC3170000, 0fC2FE0000, %p60;
|
|
selp.f32 %f379, %f377, %f108, %p60;
|
|
mov.b32 %r231, %f379;
|
|
and.b32 %r232, %r231, 8388607;
|
|
or.b32 %r233, %r232, 1065353216;
|
|
mov.b32 %f380, %r233;
|
|
shr.u32 %r234, %r231, 23;
|
|
cvt.rn.f32.u32 %f381, %r234;
|
|
add.f32 %f382, %f378, %f381;
|
|
setp.gt.f32 %p61, %f380, 0f3FB504F3;
|
|
mul.f32 %f383, %f380, 0f3F000000;
|
|
add.f32 %f384, %f382, 0f3F800000;
|
|
selp.f32 %f385, %f383, %f380, %p61;
|
|
selp.f32 %f386, %f384, %f382, %p61;
|
|
add.f32 %f387, %f385, 0fBF800000;
|
|
add.f32 %f376, %f385, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f375,%f376;
|
|
// inline asm
|
|
add.f32 %f388, %f387, %f387;
|
|
mul.f32 %f389, %f375, %f388;
|
|
mul.f32 %f390, %f389, %f389;
|
|
fma.rn.f32 %f393, %f634, %f390, %f635;
|
|
fma.rn.f32 %f395, %f393, %f390, %f636;
|
|
mul.rn.f32 %f396, %f395, %f390;
|
|
mul.rn.f32 %f397, %f396, %f389;
|
|
sub.f32 %f398, %f387, %f389;
|
|
neg.f32 %f399, %f389;
|
|
add.f32 %f400, %f398, %f398;
|
|
fma.rn.f32 %f401, %f399, %f387, %f400;
|
|
mul.rn.f32 %f402, %f375, %f401;
|
|
add.f32 %f403, %f397, %f389;
|
|
sub.f32 %f404, %f389, %f403;
|
|
add.f32 %f405, %f397, %f404;
|
|
add.f32 %f406, %f402, %f405;
|
|
add.f32 %f407, %f403, %f406;
|
|
sub.f32 %f408, %f403, %f407;
|
|
add.f32 %f409, %f406, %f408;
|
|
mul.rn.f32 %f411, %f386, %f637;
|
|
mul.rn.f32 %f413, %f386, %f638;
|
|
add.f32 %f414, %f411, %f407;
|
|
sub.f32 %f415, %f411, %f414;
|
|
add.f32 %f416, %f407, %f415;
|
|
add.f32 %f417, %f409, %f416;
|
|
add.f32 %f418, %f413, %f417;
|
|
add.f32 %f419, %f414, %f418;
|
|
sub.f32 %f420, %f414, %f419;
|
|
add.f32 %f421, %f418, %f420;
|
|
mul.rn.f32 %f423, %f633, %f419;
|
|
neg.f32 %f424, %f423;
|
|
fma.rn.f32 %f425, %f633, %f419, %f424;
|
|
fma.rn.f32 %f426, %f633, %f421, %f425;
|
|
fma.rn.f32 %f428, %f639, %f419, %f426;
|
|
add.rn.f32 %f429, %f423, %f428;
|
|
neg.f32 %f430, %f429;
|
|
add.rn.f32 %f431, %f423, %f430;
|
|
add.rn.f32 %f432, %f431, %f428;
|
|
mov.b32 %r235, %f429;
|
|
setp.eq.s32 %p62, %r235, 1118925336;
|
|
add.s32 %r236, %r235, -1;
|
|
mov.b32 %f433, %r236;
|
|
add.f32 %f434, %f432, 0f37000000;
|
|
selp.f32 %f435, %f433, %f429, %p62;
|
|
selp.f32 %f109, %f434, %f432, %p62;
|
|
mul.f32 %f436, %f435, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f437, %f436;
|
|
fma.rn.f32 %f439, %f437, %f640, %f435;
|
|
fma.rn.f32 %f441, %f437, %f641, %f439;
|
|
mul.f32 %f442, %f441, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f443, %f442;
|
|
add.f32 %f444, %f437, 0f00000000;
|
|
ex2.approx.f32 %f445, %f444;
|
|
mul.f32 %f446, %f443, %f445;
|
|
setp.lt.f32 %p63, %f435, 0fC2D20000;
|
|
selp.f32 %f447, 0f00000000, %f446, %p63;
|
|
setp.gt.f32 %p64, %f435, 0f42D20000;
|
|
selp.f32 %f702, 0f7F800000, %f447, %p64;
|
|
setp.eq.f32 %p65, %f702, 0f7F800000;
|
|
@%p65 bra BB0_70;
|
|
|
|
fma.rn.f32 %f702, %f702, %f109, %f702;
|
|
|
|
BB0_70:
|
|
setp.lt.f32 %p66, %f92, 0f00000000;
|
|
and.pred %p2, %p66, %p51;
|
|
mov.b32 %r237, %f702;
|
|
xor.b32 %r238, %r237, -2147483648;
|
|
mov.b32 %f448, %r238;
|
|
selp.f32 %f704, %f448, %f702, %p2;
|
|
setp.eq.f32 %p68, %f92, 0f00000000;
|
|
@%p68 bra BB0_73;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_73:
|
|
add.f32 %f451, %f92, %f92;
|
|
selp.f32 %f704, %f451, 0f00000000, %p51;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_71:
|
|
setp.geu.f32 %p69, %f92, 0f00000000;
|
|
@%p69 bra BB0_74;
|
|
|
|
mov.f32 %f652, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f450, %f652;
|
|
setp.neu.f32 %p70, %f450, 0f3EE66666;
|
|
selp.f32 %f704, 0f7FFFFFFF, %f704, %p70;
|
|
|
|
BB0_74:
|
|
abs.f32 %f656, %f92;
|
|
add.f32 %f452, %f656, 0f3EE66666;
|
|
mov.b32 %r239, %f452;
|
|
setp.lt.s32 %p72, %r239, 2139095040;
|
|
@%p72 bra BB0_79;
|
|
|
|
abs.f32 %f657, %f92;
|
|
setp.gtu.f32 %p73, %f657, 0f7F800000;
|
|
@%p73 bra BB0_78;
|
|
bra.uni BB0_76;
|
|
|
|
BB0_78:
|
|
add.f32 %f704, %f92, 0f3EE66666;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_76:
|
|
abs.f32 %f658, %f92;
|
|
setp.neu.f32 %p74, %f658, 0f7F800000;
|
|
@%p74 bra BB0_79;
|
|
|
|
selp.f32 %f704, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_79:
|
|
mov.f32 %f650, 0fB5BFBE8E;
|
|
mov.f32 %f649, 0fBF317200;
|
|
mov.f32 %f648, 0f00000000;
|
|
mov.f32 %f647, 0f35BFBE8E;
|
|
mov.f32 %f646, 0f3F317200;
|
|
mov.f32 %f645, 0f3DAAAABD;
|
|
mov.f32 %f644, 0f3C4CAF63;
|
|
mov.f32 %f643, 0f3B18F0FE;
|
|
mov.f32 %f642, 0f3EE66666;
|
|
setp.eq.f32 %p75, %f92, 0f3F800000;
|
|
selp.f32 %f120, 0f3F800000, %f704, %p75;
|
|
abs.f32 %f121, %f93;
|
|
setp.lt.f32 %p76, %f121, 0f00800000;
|
|
mul.f32 %f455, %f121, 0f4B800000;
|
|
selp.f32 %f456, 0fC3170000, 0fC2FE0000, %p76;
|
|
selp.f32 %f457, %f455, %f121, %p76;
|
|
mov.b32 %r240, %f457;
|
|
and.b32 %r241, %r240, 8388607;
|
|
or.b32 %r242, %r241, 1065353216;
|
|
mov.b32 %f458, %r242;
|
|
shr.u32 %r243, %r240, 23;
|
|
cvt.rn.f32.u32 %f459, %r243;
|
|
add.f32 %f460, %f456, %f459;
|
|
setp.gt.f32 %p77, %f458, 0f3FB504F3;
|
|
mul.f32 %f461, %f458, 0f3F000000;
|
|
add.f32 %f462, %f460, 0f3F800000;
|
|
selp.f32 %f463, %f461, %f458, %p77;
|
|
selp.f32 %f464, %f462, %f460, %p77;
|
|
add.f32 %f465, %f463, 0fBF800000;
|
|
add.f32 %f454, %f463, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f453,%f454;
|
|
// inline asm
|
|
add.f32 %f466, %f465, %f465;
|
|
mul.f32 %f467, %f453, %f466;
|
|
mul.f32 %f468, %f467, %f467;
|
|
fma.rn.f32 %f471, %f643, %f468, %f644;
|
|
fma.rn.f32 %f473, %f471, %f468, %f645;
|
|
mul.rn.f32 %f474, %f473, %f468;
|
|
mul.rn.f32 %f475, %f474, %f467;
|
|
sub.f32 %f476, %f465, %f467;
|
|
neg.f32 %f477, %f467;
|
|
add.f32 %f478, %f476, %f476;
|
|
fma.rn.f32 %f479, %f477, %f465, %f478;
|
|
mul.rn.f32 %f480, %f453, %f479;
|
|
add.f32 %f481, %f475, %f467;
|
|
sub.f32 %f482, %f467, %f481;
|
|
add.f32 %f483, %f475, %f482;
|
|
add.f32 %f484, %f480, %f483;
|
|
add.f32 %f485, %f481, %f484;
|
|
sub.f32 %f486, %f481, %f485;
|
|
add.f32 %f487, %f484, %f486;
|
|
mul.rn.f32 %f489, %f464, %f646;
|
|
mul.rn.f32 %f491, %f464, %f647;
|
|
add.f32 %f492, %f489, %f485;
|
|
sub.f32 %f493, %f489, %f492;
|
|
add.f32 %f494, %f485, %f493;
|
|
add.f32 %f495, %f487, %f494;
|
|
add.f32 %f496, %f491, %f495;
|
|
add.f32 %f497, %f492, %f496;
|
|
sub.f32 %f498, %f492, %f497;
|
|
add.f32 %f499, %f496, %f498;
|
|
mul.rn.f32 %f501, %f642, %f497;
|
|
neg.f32 %f502, %f501;
|
|
fma.rn.f32 %f503, %f642, %f497, %f502;
|
|
fma.rn.f32 %f504, %f642, %f499, %f503;
|
|
fma.rn.f32 %f506, %f648, %f497, %f504;
|
|
add.rn.f32 %f507, %f501, %f506;
|
|
neg.f32 %f508, %f507;
|
|
add.rn.f32 %f509, %f501, %f508;
|
|
add.rn.f32 %f510, %f509, %f506;
|
|
mov.b32 %r244, %f507;
|
|
setp.eq.s32 %p78, %r244, 1118925336;
|
|
add.s32 %r245, %r244, -1;
|
|
mov.b32 %f511, %r245;
|
|
add.f32 %f512, %f510, 0f37000000;
|
|
selp.f32 %f513, %f511, %f507, %p78;
|
|
selp.f32 %f122, %f512, %f510, %p78;
|
|
mul.f32 %f514, %f513, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f515, %f514;
|
|
fma.rn.f32 %f517, %f515, %f649, %f513;
|
|
fma.rn.f32 %f519, %f515, %f650, %f517;
|
|
mul.f32 %f520, %f519, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f521, %f520;
|
|
add.f32 %f522, %f515, 0f00000000;
|
|
ex2.approx.f32 %f523, %f522;
|
|
mul.f32 %f524, %f521, %f523;
|
|
setp.lt.f32 %p79, %f513, 0fC2D20000;
|
|
selp.f32 %f525, 0f00000000, %f524, %p79;
|
|
setp.gt.f32 %p80, %f513, 0f42D20000;
|
|
selp.f32 %f705, 0f7F800000, %f525, %p80;
|
|
setp.eq.f32 %p81, %f705, 0f7F800000;
|
|
@%p81 bra BB0_81;
|
|
|
|
fma.rn.f32 %f705, %f705, %f122, %f705;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p82, %f93, 0f00000000;
|
|
and.pred %p3, %p82, %p51;
|
|
mov.b32 %r246, %f705;
|
|
xor.b32 %r247, %r246, -2147483648;
|
|
mov.b32 %f526, %r247;
|
|
selp.f32 %f707, %f526, %f705, %p3;
|
|
setp.eq.f32 %p84, %f93, 0f00000000;
|
|
@%p84 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f529, %f93, %f93;
|
|
selp.f32 %f707, %f529, 0f00000000, %p51;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p85, %f93, 0f00000000;
|
|
@%p85 bra BB0_85;
|
|
|
|
mov.f32 %f651, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f528, %f651;
|
|
setp.neu.f32 %p86, %f528, 0f3EE66666;
|
|
selp.f32 %f707, 0f7FFFFFFF, %f707, %p86;
|
|
|
|
BB0_85:
|
|
abs.f32 %f659, %f93;
|
|
add.f32 %f530, %f659, 0f3EE66666;
|
|
mov.b32 %r248, %f530;
|
|
setp.lt.s32 %p88, %r248, 2139095040;
|
|
@%p88 bra BB0_90;
|
|
|
|
abs.f32 %f660, %f93;
|
|
setp.gtu.f32 %p89, %f660, 0f7F800000;
|
|
@%p89 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f707, %f93, 0f3EE66666;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
abs.f32 %f661, %f93;
|
|
setp.neu.f32 %p90, %f661, 0f7F800000;
|
|
@%p90 bra BB0_90;
|
|
|
|
selp.f32 %f707, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mov.u32 %r329, 4;
|
|
setp.eq.f32 %p91, %f93, 0f3F800000;
|
|
selp.f32 %f531, 0f3F800000, %f707, %p91;
|
|
cvt.u64.u32 %rd66, %r5;
|
|
cvt.u64.u32 %rd65, %r4;
|
|
mov.u64 %rd69, image;
|
|
cvta.global.u64 %rd64, %rd69;
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd64, %r92, %r329, %rd65, %rd66, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f532, %f531;
|
|
mul.f32 %f533, %f532, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r251, %f533;
|
|
cvt.sat.f32.f32 %f534, %f120;
|
|
mul.f32 %f535, %f534, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r252, %f535;
|
|
cvt.sat.f32.f32 %f536, %f107;
|
|
mul.f32 %f537, %f536, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r253, %f537;
|
|
cvt.u16.u32 %rs14, %r251;
|
|
cvt.u16.u32 %rs15, %r253;
|
|
cvt.u16.u32 %rs16, %r252;
|
|
mov.u16 %rs17, 255;
|
|
st.v4.u8 [%rd63], {%rs14, %rs16, %rs15, %rs17};
|
|
ld.global.u32 %r354, [imageEnabled];
|
|
|
|
BB0_91:
|
|
cvt.u64.u32 %rd16, %r4;
|
|
cvt.u64.u32 %rd17, %r5;
|
|
and.b32 %r254, %r354, 4;
|
|
setp.eq.s32 %p92, %r254, 0;
|
|
@%p92 bra BB0_95;
|
|
|
|
ld.global.u32 %r255, [additive];
|
|
setp.eq.s32 %p93, %r255, 0;
|
|
mov.f32 %f538, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs18, %f538;}
|
|
|
|
// inline asm
|
|
@%p93 bra BB0_94;
|
|
|
|
mov.u64 %rd82, image_HDR;
|
|
cvta.global.u64 %rd71, %rd82;
|
|
mov.u32 %r259, 8;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs25, %rs26, %rs27, %rs28}, [%rd70];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f539, %rs25;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f540, %rs26;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f541, %rs27;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd76), _rt_buffer_get_64, (%rd71, %r92, %r259, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f542, %f91, %f539;
|
|
add.f32 %f543, %f92, %f540;
|
|
add.f32 %f544, %f93, %f541;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs24, %f544;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f543;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs22, %f542;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd76], {%rs22, %rs23, %rs24, %rs18};
|
|
bra.uni BB0_95;
|
|
|
|
BB0_94:
|
|
mov.u64 %rd89, image_HDR;
|
|
cvta.global.u64 %rd84, %rd89;
|
|
mov.u32 %r261, 8;
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd84, %r92, %r261, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f93;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f92;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f91;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs29, %rs30, %rs31, %rs18};
|
|
|
|
BB0_95:
|
|
mul.f32 %f133, %f87, 0f3F000000;
|
|
ld.global.f32 %f549, [skyColor];
|
|
mul.f32 %f134, %f133, %f549;
|
|
ld.global.f32 %f550, [skyColor+4];
|
|
mul.f32 %f135, %f133, %f550;
|
|
ld.global.f32 %f551, [skyColor+8];
|
|
mul.f32 %f136, %f133, %f551;
|
|
ld.global.u32 %r262, [additive];
|
|
setp.eq.s32 %p94, %r262, 0;
|
|
mov.f32 %f548, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f548;}
|
|
|
|
// inline asm
|
|
@%p94 bra BB0_97;
|
|
|
|
mov.u64 %rd102, image_RNM0;
|
|
cvta.global.u64 %rd91, %rd102;
|
|
mov.u32 %r266, 8;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs39, %rs40, %rs41, %rs42}, [%rd90];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f552, %rs39;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f553, %rs40;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f554, %rs41;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd96), _rt_buffer_get_64, (%rd91, %r92, %r266, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f555, %f134, %f552;
|
|
add.f32 %f556, %f135, %f553;
|
|
add.f32 %f557, %f136, %f554;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f557;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f556;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f555;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd96], {%rs36, %rs37, %rs38, %rs32};
|
|
bra.uni BB0_98;
|
|
|
|
BB0_97:
|
|
mov.u64 %rd109, image_RNM0;
|
|
cvta.global.u64 %rd104, %rd109;
|
|
mov.u32 %r268, 8;
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd104, %r92, %r268, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f136;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f135;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f134;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs43, %rs44, %rs45, %rs32};
|
|
|
|
BB0_98:
|
|
mov.f32 %f562, 0f34000000;
|
|
max.f32 %f563, %f133, %f562;
|
|
mul.f32 %f564, %f88, 0f3F000000;
|
|
div.rn.f32 %f565, %f564, %f563;
|
|
fma.rn.f32 %f137, %f565, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f566, %f89, 0f3F000000;
|
|
div.rn.f32 %f567, %f566, %f563;
|
|
fma.rn.f32 %f138, %f567, 0f3F000000, 0f3F000000;
|
|
mul.f32 %f568, %f90, 0f3F000000;
|
|
div.rn.f32 %f569, %f568, %f563;
|
|
fma.rn.f32 %f139, %f569, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r269, [additive];
|
|
setp.eq.s32 %p95, %r269, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f548;}
|
|
|
|
// inline asm
|
|
@%p95 bra BB0_100;
|
|
|
|
mov.u64 %rd122, image_RNM1;
|
|
cvta.global.u64 %rd111, %rd122;
|
|
mov.u32 %r273, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs53, %rs54, %rs55, %rs56}, [%rd110];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f570, %rs53;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f571, %rs54;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f572, %rs55;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd116), _rt_buffer_get_64, (%rd111, %r92, %r273, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f573, %f137, %f570;
|
|
add.f32 %f574, %f137, %f571;
|
|
add.f32 %f575, %f137, %f572;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f575;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f574;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f573;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd116], {%rs50, %rs51, %rs52, %rs46};
|
|
bra.uni BB0_101;
|
|
|
|
BB0_100:
|
|
mov.u64 %rd129, image_RNM1;
|
|
cvta.global.u64 %rd124, %rd129;
|
|
mov.u32 %r275, 8;
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd124, %r92, %r275, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f137;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs57, %rs57, %rs57, %rs46};
|
|
|
|
BB0_101:
|
|
ld.global.u32 %r276, [additive];
|
|
setp.eq.s32 %p96, %r276, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f548;}
|
|
|
|
// inline asm
|
|
@%p96 bra BB0_103;
|
|
|
|
mov.u64 %rd142, image_RNM2;
|
|
cvta.global.u64 %rd131, %rd142;
|
|
mov.u32 %r280, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs65, %rs66, %rs67, %rs68}, [%rd130];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f578, %rs65;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f579, %rs66;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f580, %rs67;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd136), _rt_buffer_get_64, (%rd131, %r92, %r280, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f581, %f138, %f578;
|
|
add.f32 %f582, %f138, %f579;
|
|
add.f32 %f583, %f138, %f580;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f583;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f582;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f581;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd136], {%rs62, %rs63, %rs64, %rs58};
|
|
bra.uni BB0_104;
|
|
|
|
BB0_103:
|
|
mov.u64 %rd149, image_RNM2;
|
|
cvta.global.u64 %rd144, %rd149;
|
|
mov.u32 %r282, 8;
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd144, %r92, %r282, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f138;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs69, %rs69, %rs69, %rs58};
|
|
|
|
BB0_104:
|
|
ld.global.u32 %r283, [additive];
|
|
setp.eq.s32 %p97, %r283, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f548;}
|
|
|
|
// inline asm
|
|
@%p97 bra BB0_106;
|
|
|
|
mov.u64 %rd162, image_RNM3;
|
|
cvta.global.u64 %rd151, %rd162;
|
|
mov.u32 %r287, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs77, %rs78, %rs79, %rs80}, [%rd150];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f586, %rs77;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f587, %rs78;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f588, %rs79;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd156), _rt_buffer_get_64, (%rd151, %r92, %r287, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f589, %f139, %f586;
|
|
add.f32 %f590, %f139, %f587;
|
|
add.f32 %f591, %f139, %f588;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f591;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f590;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f589;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd156], {%rs74, %rs75, %rs76, %rs70};
|
|
bra.uni BB0_125;
|
|
|
|
BB0_106:
|
|
mov.u64 %rd169, image_RNM3;
|
|
cvta.global.u64 %rd164, %rd169;
|
|
mov.u32 %r289, 8;
|
|
// inline asm
|
|
call (%rd163), _rt_buffer_get_64, (%rd164, %r92, %r289, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f139;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd163], {%rs81, %rs81, %rs81, %rs70};
|
|
|
|
BB0_125:
|
|
ret;
|
|
}
|
|
|
|
|