2595 lines
74 KiB
Plaintext
2595 lines
74 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_Mask[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 uvtangent[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 1 .b8 rnd_seeds2[1];
|
|
.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_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[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 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 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 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_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 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 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.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_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[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[44];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<157>;
|
|
.reg .b16 %rs<158>;
|
|
.reg .f32 %f<1384>;
|
|
.reg .b32 %r<432>;
|
|
.reg .b64 %rd<314>;
|
|
|
|
|
|
mov.u64 %rd313, __local_depot0;
|
|
cvta.local.u64 %SP, %rd313;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r111, %r112}, [pixelID];
|
|
cvt.u64.u32 %rd27, %r111;
|
|
cvt.u64.u32 %rd28, %r112;
|
|
mov.u64 %rd31, uvnormal;
|
|
cvta.global.u64 %rd26, %rd31;
|
|
mov.u32 %r109, 2;
|
|
mov.u32 %r110, 4;
|
|
mov.u64 %rd30, 0;
|
|
// inline asm
|
|
call (%rd25), _rt_buffer_get_64, (%rd26, %r109, %r110, %rd27, %rd28, %rd30, %rd30);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd25];
|
|
shr.u32 %r115, %r4, 16;
|
|
cvt.u16.u32 %rs1, %r115;
|
|
and.b16 %rs6, %rs1, 255;
|
|
cvt.u16.u32 %rs7, %r4;
|
|
or.b16 %rs8, %rs7, %rs6;
|
|
setp.eq.s16 %p7, %rs8, 0;
|
|
mov.f32 %f1244, 0f00000000;
|
|
mov.f32 %f1245, %f1244;
|
|
mov.f32 %f1246, %f1244;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs9, [%rd25+1];
|
|
and.b16 %rs11, %rs7, 255;
|
|
cvt.rn.f32.u16 %f340, %rs11;
|
|
div.rn.f32 %f341, %f340, 0f437F0000;
|
|
fma.rn.f32 %f342, %f341, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f343, %rs9;
|
|
div.rn.f32 %f344, %f343, 0f437F0000;
|
|
fma.rn.f32 %f345, %f344, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f346, %rs6;
|
|
div.rn.f32 %f347, %f346, 0f437F0000;
|
|
fma.rn.f32 %f348, %f347, 0f40000000, 0fBF800000;
|
|
mul.f32 %f349, %f345, %f345;
|
|
fma.rn.f32 %f350, %f342, %f342, %f349;
|
|
fma.rn.f32 %f351, %f348, %f348, %f350;
|
|
sqrt.rn.f32 %f352, %f351;
|
|
rcp.rn.f32 %f353, %f352;
|
|
mul.f32 %f1244, %f342, %f353;
|
|
mul.f32 %f1245, %f345, %f353;
|
|
mul.f32 %f1246, %f348, %f353;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r116, %r117}, [pixelID];
|
|
ld.global.v2.u32 {%r119, %r120}, [tileInfo];
|
|
add.s32 %r5, %r116, %r119;
|
|
add.s32 %r6, %r117, %r120;
|
|
setp.eq.f32 %p8, %f1245, 0f00000000;
|
|
setp.eq.f32 %p9, %f1244, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1246, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_145;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_145:
|
|
ld.global.u32 %r431, [imageEnabled];
|
|
and.b32 %r360, %r431, 1;
|
|
setp.eq.b32 %p149, %r360, 1;
|
|
@!%p149 bra BB0_147;
|
|
bra.uni BB0_146;
|
|
|
|
BB0_146:
|
|
cvt.u64.u32 %rd210, %r5;
|
|
cvt.u64.u32 %rd211, %r6;
|
|
mov.u64 %rd214, image;
|
|
cvta.global.u64 %rd209, %rd214;
|
|
mov.u64 %rd213, 0;
|
|
// inline asm
|
|
call (%rd208), _rt_buffer_get_64, (%rd209, %r109, %r110, %rd210, %rd211, %rd213, %rd213);
|
|
// inline asm
|
|
mov.u16 %rs103, 0;
|
|
st.v4.u8 [%rd208], {%rs103, %rs103, %rs103, %rs103};
|
|
ld.global.u32 %r431, [imageEnabled];
|
|
|
|
BB0_147:
|
|
and.b32 %r363, %r431, 8;
|
|
setp.eq.s32 %p150, %r363, 0;
|
|
@%p150 bra BB0_149;
|
|
|
|
cvt.u64.u32 %rd218, %r6;
|
|
cvt.u64.u32 %rd217, %r5;
|
|
mov.u64 %rd221, image_Mask;
|
|
cvta.global.u64 %rd216, %rd221;
|
|
mov.u64 %rd220, 0;
|
|
// inline asm
|
|
call (%rd215), _rt_buffer_get_64, (%rd216, %r109, %r109, %rd217, %rd218, %rd220, %rd220);
|
|
// inline asm
|
|
mov.f32 %f1206, 0f00000000;
|
|
cvt.rzi.u32.f32 %r366, %f1206;
|
|
cvt.u16.u32 %rs104, %r366;
|
|
mov.u16 %rs105, 0;
|
|
st.v2.u8 [%rd215], {%rs104, %rs105};
|
|
ld.global.u32 %r431, [imageEnabled];
|
|
|
|
BB0_149:
|
|
cvt.u64.u32 %rd23, %r5;
|
|
cvt.u64.u32 %rd24, %r6;
|
|
and.b32 %r367, %r431, 4;
|
|
setp.eq.s32 %p151, %r367, 0;
|
|
@%p151 bra BB0_153;
|
|
|
|
ld.global.u32 %r368, [additive];
|
|
setp.eq.s32 %p152, %r368, 0;
|
|
@%p152 bra BB0_152;
|
|
|
|
mov.u64 %rd234, image_HDR;
|
|
cvta.global.u64 %rd223, %rd234;
|
|
mov.u32 %r372, 8;
|
|
mov.u64 %rd233, 0;
|
|
// inline asm
|
|
call (%rd222), _rt_buffer_get_64, (%rd223, %r109, %r372, %rd23, %rd24, %rd233, %rd233);
|
|
// inline asm
|
|
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd222];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1207, %rs112;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1208, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1209, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd228), _rt_buffer_get_64, (%rd223, %r109, %r372, %rd23, %rd24, %rd233, %rd233);
|
|
// inline asm
|
|
add.f32 %f1210, %f1207, 0f00000000;
|
|
add.f32 %f1211, %f1208, 0f00000000;
|
|
add.f32 %f1212, %f1209, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1212;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1211;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f1210;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs116, 0;
|
|
st.v4.u16 [%rd228], {%rs109, %rs110, %rs111, %rs116};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r128, %r129}, [pixelID];
|
|
cvt.u64.u32 %rd34, %r128;
|
|
cvt.u64.u32 %rd35, %r129;
|
|
mov.u64 %rd44, uvpos;
|
|
cvta.global.u64 %rd33, %rd44;
|
|
mov.u32 %r125, 12;
|
|
// inline asm
|
|
call (%rd32), _rt_buffer_get_64, (%rd33, %r109, %r125, %rd34, %rd35, %rd30, %rd30);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd32+8];
|
|
ld.f32 %f8, [%rd32+4];
|
|
ld.f32 %f7, [%rd32];
|
|
mul.f32 %f357, %f7, 0f3456BF95;
|
|
mul.f32 %f358, %f8, 0f3456BF95;
|
|
mul.f32 %f359, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f1244;
|
|
div.rn.f32 %f360, %f357, %f10;
|
|
abs.f32 %f361, %f1245;
|
|
div.rn.f32 %f362, %f358, %f361;
|
|
abs.f32 %f11, %f1246;
|
|
div.rn.f32 %f363, %f359, %f11;
|
|
abs.f32 %f364, %f360;
|
|
abs.f32 %f365, %f362;
|
|
abs.f32 %f366, %f363;
|
|
mov.f32 %f367, 0f38D1B717;
|
|
max.f32 %f368, %f364, %f367;
|
|
max.f32 %f369, %f365, %f367;
|
|
max.f32 %f370, %f366, %f367;
|
|
fma.rn.f32 %f12, %f1244, %f368, %f7;
|
|
fma.rn.f32 %f13, %f1245, %f369, %f8;
|
|
fma.rn.f32 %f14, %f1246, %f370, %f9;
|
|
ld.global.v2.u32 {%r132, %r133}, [pixelID];
|
|
cvt.u64.u32 %rd40, %r132;
|
|
cvt.u64.u32 %rd41, %r133;
|
|
mov.u64 %rd45, uvtangent;
|
|
cvta.global.u64 %rd39, %rd45;
|
|
// inline asm
|
|
call (%rd38), _rt_buffer_get_64, (%rd39, %r109, %r110, %rd40, %rd41, %rd30, %rd30);
|
|
// inline asm
|
|
ld.u32 %r7, [%rd38];
|
|
shr.u32 %r8, %r7, 16;
|
|
cvt.u16.u32 %rs13, %r8;
|
|
and.b16 %rs14, %rs13, 255;
|
|
cvt.u16.u32 %rs15, %r7;
|
|
or.b16 %rs16, %rs15, %rs14;
|
|
setp.eq.s16 %p13, %rs16, 0;
|
|
mov.f32 %f49, 0f00000000;
|
|
mov.f32 %f1247, %f49;
|
|
mov.f32 %f1248, %f49;
|
|
mov.f32 %f1249, %f49;
|
|
@%p13 bra BB0_5;
|
|
|
|
ld.u8 %rs17, [%rd38+1];
|
|
and.b16 %rs19, %rs15, 255;
|
|
cvt.rn.f32.u16 %f371, %rs19;
|
|
div.rn.f32 %f372, %f371, 0f437F0000;
|
|
fma.rn.f32 %f373, %f372, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f374, %rs17;
|
|
div.rn.f32 %f375, %f374, 0f437F0000;
|
|
fma.rn.f32 %f376, %f375, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f377, %rs14;
|
|
div.rn.f32 %f378, %f377, 0f437F0000;
|
|
fma.rn.f32 %f379, %f378, 0f40000000, 0fBF800000;
|
|
mul.f32 %f380, %f376, %f376;
|
|
fma.rn.f32 %f381, %f373, %f373, %f380;
|
|
fma.rn.f32 %f382, %f379, %f379, %f381;
|
|
sqrt.rn.f32 %f383, %f382;
|
|
rcp.rn.f32 %f384, %f383;
|
|
mul.f32 %f1247, %f373, %f384;
|
|
mul.f32 %f1248, %f376, %f384;
|
|
mul.f32 %f1249, %f379, %f384;
|
|
|
|
BB0_5:
|
|
mul.f32 %f388, %f1246, %f1248;
|
|
mul.f32 %f389, %f1245, %f1249;
|
|
sub.f32 %f390, %f389, %f388;
|
|
mul.f32 %f391, %f1244, %f1249;
|
|
mul.f32 %f392, %f1246, %f1247;
|
|
sub.f32 %f393, %f392, %f391;
|
|
mul.f32 %f394, %f1245, %f1247;
|
|
mul.f32 %f395, %f1244, %f1248;
|
|
sub.f32 %f396, %f395, %f394;
|
|
setp.lt.u32 %p14, %r7, 16777216;
|
|
selp.f32 %f397, 0fBF800000, 0f3F800000, %p14;
|
|
mul.f32 %f398, %f390, %f397;
|
|
mul.f32 %f399, %f393, %f397;
|
|
mul.f32 %f400, %f396, %f397;
|
|
mul.f32 %f401, %f398, 0f00000000;
|
|
mul.f32 %f402, %f399, 0f00000000;
|
|
mul.f32 %f403, %f400, 0f00000000;
|
|
fma.rn.f32 %f404, %f1247, 0f3F5105EC, %f401;
|
|
fma.rn.f32 %f405, %f1248, 0f3F5105EC, %f402;
|
|
fma.rn.f32 %f406, %f1249, 0f3F5105EC, %f403;
|
|
mul.f32 %f21, %f1244, 0f3F13CD3A;
|
|
add.f32 %f22, %f21, %f404;
|
|
mul.f32 %f23, %f1245, 0f3F13CD3A;
|
|
add.f32 %f24, %f23, %f405;
|
|
mul.f32 %f25, %f1246, 0f3F13CD3A;
|
|
add.f32 %f26, %f25, %f406;
|
|
ld.global.v2.u32 {%r138, %r139}, [pixelID];
|
|
cvt.u64.u32 %rd48, %r138;
|
|
cvt.u64.u32 %rd49, %r139;
|
|
// inline asm
|
|
call (%rd46), _rt_buffer_get_64, (%rd39, %r109, %r110, %rd48, %rd49, %rd30, %rd30);
|
|
// inline asm
|
|
ld.u32 %r9, [%rd46];
|
|
shr.u32 %r10, %r9, 16;
|
|
cvt.u16.u32 %rs22, %r10;
|
|
and.b16 %rs23, %rs22, 255;
|
|
cvt.u16.u32 %rs24, %r9;
|
|
or.b16 %rs25, %rs24, %rs23;
|
|
setp.eq.s16 %p15, %rs25, 0;
|
|
mov.f32 %f1250, %f49;
|
|
mov.f32 %f1251, %f49;
|
|
mov.f32 %f1252, %f49;
|
|
@%p15 bra BB0_7;
|
|
|
|
ld.u8 %rs26, [%rd46+1];
|
|
and.b16 %rs28, %rs24, 255;
|
|
cvt.rn.f32.u16 %f407, %rs28;
|
|
div.rn.f32 %f408, %f407, 0f437F0000;
|
|
fma.rn.f32 %f409, %f408, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f410, %rs26;
|
|
div.rn.f32 %f411, %f410, 0f437F0000;
|
|
fma.rn.f32 %f412, %f411, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f413, %rs23;
|
|
div.rn.f32 %f414, %f413, 0f437F0000;
|
|
fma.rn.f32 %f415, %f414, 0f40000000, 0fBF800000;
|
|
mul.f32 %f416, %f412, %f412;
|
|
fma.rn.f32 %f417, %f409, %f409, %f416;
|
|
fma.rn.f32 %f418, %f415, %f415, %f417;
|
|
sqrt.rn.f32 %f419, %f418;
|
|
rcp.rn.f32 %f420, %f419;
|
|
mul.f32 %f1250, %f409, %f420;
|
|
mul.f32 %f1251, %f412, %f420;
|
|
mul.f32 %f1252, %f415, %f420;
|
|
|
|
BB0_7:
|
|
mul.f32 %f424, %f1246, %f1251;
|
|
mul.f32 %f425, %f1245, %f1252;
|
|
sub.f32 %f426, %f425, %f424;
|
|
mul.f32 %f427, %f1244, %f1252;
|
|
mul.f32 %f428, %f1246, %f1250;
|
|
sub.f32 %f429, %f428, %f427;
|
|
mul.f32 %f430, %f1245, %f1250;
|
|
mul.f32 %f431, %f1244, %f1251;
|
|
sub.f32 %f432, %f431, %f430;
|
|
setp.lt.u32 %p16, %r9, 16777216;
|
|
selp.f32 %f433, 0fBF800000, 0f3F800000, %p16;
|
|
mul.f32 %f434, %f426, %f433;
|
|
mul.f32 %f435, %f429, %f433;
|
|
mul.f32 %f436, %f432, %f433;
|
|
mul.f32 %f437, %f434, 0f3F3504F3;
|
|
mul.f32 %f438, %f435, 0f3F3504F3;
|
|
mul.f32 %f439, %f436, 0f3F3504F3;
|
|
fma.rn.f32 %f440, %f1250, 0fBED105EC, %f437;
|
|
fma.rn.f32 %f441, %f1251, 0fBED105EC, %f438;
|
|
fma.rn.f32 %f442, %f1252, 0fBED105EC, %f439;
|
|
add.f32 %f33, %f21, %f440;
|
|
add.f32 %f34, %f23, %f441;
|
|
add.f32 %f35, %f25, %f442;
|
|
ld.global.v2.u32 {%r144, %r145}, [pixelID];
|
|
cvt.u64.u32 %rd55, %r144;
|
|
cvt.u64.u32 %rd56, %r145;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd39, %r109, %r110, %rd55, %rd56, %rd30, %rd30);
|
|
// inline asm
|
|
ld.u32 %r11, [%rd53];
|
|
shr.u32 %r12, %r11, 16;
|
|
cvt.u16.u32 %rs31, %r12;
|
|
and.b16 %rs32, %rs31, 255;
|
|
cvt.u16.u32 %rs33, %r11;
|
|
or.b16 %rs34, %rs33, %rs32;
|
|
setp.eq.s16 %p17, %rs34, 0;
|
|
mov.f32 %f1253, %f49;
|
|
mov.f32 %f1254, %f49;
|
|
mov.f32 %f1255, %f49;
|
|
@%p17 bra BB0_9;
|
|
|
|
ld.u8 %rs35, [%rd53+1];
|
|
and.b16 %rs37, %rs33, 255;
|
|
cvt.rn.f32.u16 %f443, %rs37;
|
|
div.rn.f32 %f444, %f443, 0f437F0000;
|
|
fma.rn.f32 %f445, %f444, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f446, %rs35;
|
|
div.rn.f32 %f447, %f446, 0f437F0000;
|
|
fma.rn.f32 %f448, %f447, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f449, %rs32;
|
|
div.rn.f32 %f450, %f449, 0f437F0000;
|
|
fma.rn.f32 %f451, %f450, 0f40000000, 0fBF800000;
|
|
mul.f32 %f452, %f448, %f448;
|
|
fma.rn.f32 %f453, %f445, %f445, %f452;
|
|
fma.rn.f32 %f454, %f451, %f451, %f453;
|
|
sqrt.rn.f32 %f455, %f454;
|
|
rcp.rn.f32 %f456, %f455;
|
|
mul.f32 %f1253, %f445, %f456;
|
|
mul.f32 %f1254, %f448, %f456;
|
|
mul.f32 %f1255, %f451, %f456;
|
|
|
|
BB0_9:
|
|
mul.f32 %f470, %f1246, %f1254;
|
|
mul.f32 %f471, %f1245, %f1255;
|
|
sub.f32 %f472, %f471, %f470;
|
|
mul.f32 %f473, %f1244, %f1255;
|
|
mul.f32 %f474, %f1246, %f1253;
|
|
sub.f32 %f475, %f474, %f473;
|
|
mul.f32 %f476, %f1245, %f1253;
|
|
mul.f32 %f477, %f1244, %f1254;
|
|
sub.f32 %f478, %f477, %f476;
|
|
setp.lt.u32 %p19, %r11, 16777216;
|
|
selp.f32 %f479, 0fBF800000, 0f3F800000, %p19;
|
|
mul.f32 %f480, %f472, %f479;
|
|
mul.f32 %f481, %f475, %f479;
|
|
mul.f32 %f482, %f478, %f479;
|
|
mul.f32 %f483, %f480, 0fBF3504F3;
|
|
mul.f32 %f484, %f481, 0fBF3504F3;
|
|
mul.f32 %f485, %f482, 0fBF3504F3;
|
|
fma.rn.f32 %f486, %f1253, 0fBED105EC, %f483;
|
|
fma.rn.f32 %f487, %f1254, 0fBED105EC, %f484;
|
|
fma.rn.f32 %f488, %f1255, 0fBED105EC, %f485;
|
|
add.f32 %f42, %f21, %f486;
|
|
add.f32 %f43, %f23, %f487;
|
|
add.f32 %f44, %f25, %f488;
|
|
ld.global.v2.u32 {%r151, %r152}, [pixelID];
|
|
cvt.u64.u32 %rd62, %r151;
|
|
cvt.u64.u32 %rd63, %r152;
|
|
mov.u64 %rd66, rnd_seeds;
|
|
cvta.global.u64 %rd61, %rd66;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd61, %r109, %r110, %rd62, %rd63, %rd30, %rd30);
|
|
// inline asm
|
|
ld.u32 %r407, [%rd60];
|
|
mov.pred %p18, 0;
|
|
mov.u32 %r15, 0;
|
|
setp.lt.s32 %p20, %r2, 1;
|
|
@%p20 bra BB0_10;
|
|
|
|
ld.global.f32 %f45, [lightPointSize];
|
|
mul.f32 %f46, %f12, 0f3456BF95;
|
|
mul.f32 %f47, %f13, 0f3456BF95;
|
|
mul.f32 %f48, %f14, 0f3456BF95;
|
|
mov.f32 %f49, 0f00000000;
|
|
mov.u32 %r400, 0;
|
|
abs.f32 %f625, %f47;
|
|
abs.f32 %f626, %f46;
|
|
max.f32 %f627, %f626, %f625;
|
|
abs.f32 %f628, %f48;
|
|
max.f32 %f629, %f627, %f628;
|
|
mov.u32 %r15, %r400;
|
|
mov.f32 %f50, %f49;
|
|
mov.f32 %f51, %f49;
|
|
mov.f32 %f52, %f49;
|
|
mov.f32 %f53, %f49;
|
|
mov.f32 %f54, %f49;
|
|
mov.f32 %f55, %f49;
|
|
mov.f32 %f56, %f49;
|
|
mov.f32 %f57, %f49;
|
|
mov.f32 %f58, %f49;
|
|
mov.f32 %f59, %f49;
|
|
mov.f32 %f60, %f49;
|
|
mov.f32 %f61, %f49;
|
|
|
|
BB0_12:
|
|
mul.lo.s32 %r16, %r400, 3;
|
|
cvt.s64.s32 %rd69, %r16;
|
|
mov.u64 %rd73, lightMeshBuffer;
|
|
cvta.global.u64 %rd68, %rd73;
|
|
mov.u32 %r157, 1;
|
|
mov.u64 %rd72, 0;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd68, %r157, %r125, %rd69, %rd72, %rd72, %rd72);
|
|
// inline asm
|
|
ld.f32 %f502, [%rd67];
|
|
sub.f32 %f503, %f502, %f7;
|
|
ld.f32 %f504, [%rd67+4];
|
|
sub.f32 %f505, %f504, %f8;
|
|
ld.f32 %f506, [%rd67+8];
|
|
sub.f32 %f507, %f506, %f9;
|
|
mul.f32 %f508, %f505, %f505;
|
|
fma.rn.f32 %f509, %f503, %f503, %f508;
|
|
fma.rn.f32 %f510, %f507, %f507, %f509;
|
|
sqrt.rn.f32 %f62, %f510;
|
|
rcp.rn.f32 %f511, %f62;
|
|
mul.f32 %f63, %f503, %f511;
|
|
mul.f32 %f64, %f505, %f511;
|
|
mul.f32 %f65, %f507, %f511;
|
|
mul.f32 %f512, %f1245, %f64;
|
|
fma.rn.f32 %f513, %f1244, %f63, %f512;
|
|
fma.rn.f32 %f66, %f1246, %f65, %f513;
|
|
setp.leu.f32 %p21, %f66, 0f00000000;
|
|
@%p21 bra BB0_29;
|
|
|
|
mov.u64 %rd305, 0;
|
|
setp.eq.s32 %p22, %r3, 0;
|
|
mul.f32 %f514, %f62, %f62;
|
|
div.rn.f32 %f67, %f45, %f514;
|
|
ld.global.v2.u32 {%r161, %r162}, [pixelID];
|
|
cvt.u64.u32 %rd80, %r162;
|
|
cvt.u64.u32 %rd81, %r161;
|
|
and.b64 %rd76, %rd81, 255;
|
|
and.b64 %rd77, %rd80, 255;
|
|
mov.u64 %rd82, rnd_seeds2;
|
|
cvta.global.u64 %rd75, %rd82;
|
|
mov.u32 %r160, 3;
|
|
// inline asm
|
|
call (%rd74), _rt_buffer_get_64, (%rd75, %r109, %r160, %rd76, %rd77, %rd305, %rd305);
|
|
// inline asm
|
|
@%p22 bra BB0_15;
|
|
|
|
ld.u8 %rs40, [%rd74];
|
|
cvt.rn.f32.u16 %f515, %rs40;
|
|
div.rn.f32 %f516, %f515, 0fC37F0000;
|
|
fma.rn.f32 %f517, %f516, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p24, %f67, %f517;
|
|
mov.pred %p156, -1;
|
|
@%p24 bra BB0_31;
|
|
|
|
BB0_15:
|
|
ld.global.f32 %f520, [lightInvCutoff];
|
|
mul.f32 %f68, %f62, %f520;
|
|
mov.f32 %f524, 0f40800000;
|
|
abs.f32 %f70, %f68;
|
|
setp.lt.f32 %p25, %f70, 0f00800000;
|
|
mul.f32 %f526, %f70, 0f4B800000;
|
|
selp.f32 %f527, 0fC3170000, 0fC2FE0000, %p25;
|
|
selp.f32 %f528, %f526, %f70, %p25;
|
|
mov.b32 %r165, %f528;
|
|
and.b32 %r166, %r165, 8388607;
|
|
or.b32 %r167, %r166, 1065353216;
|
|
mov.b32 %f529, %r167;
|
|
shr.u32 %r168, %r165, 23;
|
|
cvt.rn.f32.u32 %f530, %r168;
|
|
add.f32 %f531, %f527, %f530;
|
|
setp.gt.f32 %p26, %f529, 0f3FB504F3;
|
|
mul.f32 %f532, %f529, 0f3F000000;
|
|
add.f32 %f533, %f531, 0f3F800000;
|
|
selp.f32 %f534, %f532, %f529, %p26;
|
|
selp.f32 %f535, %f533, %f531, %p26;
|
|
add.f32 %f536, %f534, 0fBF800000;
|
|
add.f32 %f519, %f534, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f518,%f519;
|
|
// inline asm
|
|
add.f32 %f537, %f536, %f536;
|
|
mul.f32 %f538, %f518, %f537;
|
|
mul.f32 %f539, %f538, %f538;
|
|
mov.f32 %f540, 0f3C4CAF63;
|
|
mov.f32 %f541, 0f3B18F0FE;
|
|
fma.rn.f32 %f542, %f541, %f539, %f540;
|
|
mov.f32 %f543, 0f3DAAAABD;
|
|
fma.rn.f32 %f544, %f542, %f539, %f543;
|
|
mul.rn.f32 %f545, %f544, %f539;
|
|
mul.rn.f32 %f546, %f545, %f538;
|
|
sub.f32 %f547, %f536, %f538;
|
|
neg.f32 %f548, %f538;
|
|
add.f32 %f549, %f547, %f547;
|
|
fma.rn.f32 %f550, %f548, %f536, %f549;
|
|
mul.rn.f32 %f551, %f518, %f550;
|
|
add.f32 %f552, %f546, %f538;
|
|
sub.f32 %f553, %f538, %f552;
|
|
add.f32 %f554, %f546, %f553;
|
|
add.f32 %f555, %f551, %f554;
|
|
add.f32 %f556, %f552, %f555;
|
|
sub.f32 %f557, %f552, %f556;
|
|
add.f32 %f558, %f555, %f557;
|
|
mov.f32 %f559, 0f3F317200;
|
|
mul.rn.f32 %f560, %f535, %f559;
|
|
mov.f32 %f561, 0f35BFBE8E;
|
|
mul.rn.f32 %f562, %f535, %f561;
|
|
add.f32 %f563, %f560, %f556;
|
|
sub.f32 %f564, %f560, %f563;
|
|
add.f32 %f565, %f556, %f564;
|
|
add.f32 %f566, %f558, %f565;
|
|
add.f32 %f567, %f562, %f566;
|
|
add.f32 %f568, %f563, %f567;
|
|
sub.f32 %f569, %f563, %f568;
|
|
add.f32 %f570, %f567, %f569;
|
|
mul.rn.f32 %f571, %f524, %f568;
|
|
neg.f32 %f572, %f571;
|
|
fma.rn.f32 %f573, %f524, %f568, %f572;
|
|
fma.rn.f32 %f574, %f524, %f570, %f573;
|
|
mov.f32 %f575, 0f00000000;
|
|
fma.rn.f32 %f576, %f575, %f568, %f574;
|
|
add.rn.f32 %f577, %f571, %f576;
|
|
neg.f32 %f578, %f577;
|
|
add.rn.f32 %f579, %f571, %f578;
|
|
add.rn.f32 %f580, %f579, %f576;
|
|
mov.b32 %r169, %f577;
|
|
setp.eq.s32 %p27, %r169, 1118925336;
|
|
add.s32 %r170, %r169, -1;
|
|
mov.b32 %f581, %r170;
|
|
add.f32 %f582, %f580, 0f37000000;
|
|
selp.f32 %f583, %f581, %f577, %p27;
|
|
selp.f32 %f71, %f582, %f580, %p27;
|
|
mul.f32 %f584, %f583, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f585, %f584;
|
|
mov.f32 %f586, 0fBF317200;
|
|
fma.rn.f32 %f587, %f585, %f586, %f583;
|
|
mov.f32 %f588, 0fB5BFBE8E;
|
|
fma.rn.f32 %f589, %f585, %f588, %f587;
|
|
mul.f32 %f590, %f589, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f591, %f590;
|
|
add.f32 %f592, %f585, 0f00000000;
|
|
ex2.approx.f32 %f593, %f592;
|
|
mul.f32 %f594, %f591, %f593;
|
|
setp.lt.f32 %p28, %f583, 0fC2D20000;
|
|
selp.f32 %f595, 0f00000000, %f594, %p28;
|
|
setp.gt.f32 %p29, %f583, 0f42D20000;
|
|
selp.f32 %f1269, 0f7F800000, %f595, %p29;
|
|
setp.eq.f32 %p30, %f1269, 0f7F800000;
|
|
@%p30 bra BB0_17;
|
|
|
|
fma.rn.f32 %f1269, %f1269, %f71, %f1269;
|
|
|
|
BB0_17:
|
|
mov.f32 %f1240, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1239, %f1240;
|
|
add.f32 %f1238, %f1239, %f1239;
|
|
mov.f32 %f1237, 0f40800000;
|
|
sub.f32 %f1236, %f1237, %f1238;
|
|
abs.f32 %f1235, %f1236;
|
|
setp.lt.f32 %p31, %f68, 0f00000000;
|
|
setp.eq.f32 %p32, %f1235, 0f3F800000;
|
|
and.pred %p1, %p31, %p32;
|
|
mov.b32 %r171, %f1269;
|
|
xor.b32 %r172, %r171, -2147483648;
|
|
mov.b32 %f596, %r172;
|
|
selp.f32 %f1271, %f596, %f1269, %p1;
|
|
setp.eq.f32 %p33, %f68, 0f00000000;
|
|
@%p33 bra BB0_20;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_20:
|
|
add.f32 %f599, %f68, %f68;
|
|
selp.f32 %f1271, %f599, 0f00000000, %p32;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_18:
|
|
setp.geu.f32 %p34, %f68, 0f00000000;
|
|
@%p34 bra BB0_21;
|
|
|
|
mov.f32 %f1243, 0f40800000;
|
|
cvt.rzi.f32.f32 %f598, %f1243;
|
|
setp.neu.f32 %p35, %f598, 0f40800000;
|
|
selp.f32 %f1271, 0f7FFFFFFF, %f1271, %p35;
|
|
|
|
BB0_21:
|
|
add.f32 %f600, %f70, 0f40800000;
|
|
mov.b32 %r173, %f600;
|
|
setp.lt.s32 %p37, %r173, 2139095040;
|
|
@%p37 bra BB0_26;
|
|
|
|
setp.gtu.f32 %p38, %f70, 0f7F800000;
|
|
@%p38 bra BB0_25;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_25:
|
|
add.f32 %f1271, %f68, 0f40800000;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_23:
|
|
setp.neu.f32 %p39, %f70, 0f7F800000;
|
|
@%p39 bra BB0_26;
|
|
|
|
selp.f32 %f1271, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_26:
|
|
mov.u32 %r397, 1;
|
|
mov.u64 %rd304, lightMeshBuffer;
|
|
cvta.global.u64 %rd303, %rd304;
|
|
mul.lo.s32 %r396, %r400, 3;
|
|
mov.u64 %rd302, 0;
|
|
mov.f32 %f601, 0f3F800000;
|
|
sub.f32 %f602, %f601, %f1271;
|
|
setp.eq.f32 %p40, %f68, 0f3F800000;
|
|
selp.f32 %f603, 0f00000000, %f602, %p40;
|
|
cvt.sat.f32.f32 %f604, %f603;
|
|
mul.f32 %f605, %f67, %f604;
|
|
add.s32 %r176, %r396, 1;
|
|
cvt.s64.s32 %rd85, %r176;
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd303, %r397, %r125, %rd85, %rd302, %rd302, %rd302);
|
|
// inline asm
|
|
ld.f32 %f606, [%rd83];
|
|
mul.f32 %f607, %f63, %f606;
|
|
ld.f32 %f608, [%rd83+4];
|
|
mul.f32 %f609, %f64, %f608;
|
|
neg.f32 %f610, %f609;
|
|
sub.f32 %f611, %f610, %f607;
|
|
ld.f32 %f612, [%rd83+8];
|
|
mul.f32 %f613, %f65, %f612;
|
|
sub.f32 %f614, %f611, %f613;
|
|
cvt.sat.f32.f32 %f615, %f614;
|
|
mul.f32 %f82, %f605, %f615;
|
|
setp.leu.f32 %p41, %f82, 0f3727C5AC;
|
|
@%p41 bra BB0_28;
|
|
|
|
mov.u32 %r399, 1;
|
|
mov.u64 %rd308, lightMeshBuffer;
|
|
cvta.global.u64 %rd307, %rd308;
|
|
mul.lo.s32 %r398, %r400, 3;
|
|
mov.u64 %rd306, 0;
|
|
cvt.sat.f32.f32 %f624, %f66;
|
|
add.u64 %rd90, %SP, 12;
|
|
cvta.to.local.u64 %rd97, %rd90;
|
|
max.f32 %f622, %f629, %f367;
|
|
sub.f32 %f623, %f62, %f622;
|
|
mov.u32 %r182, 1065353216;
|
|
st.local.u32 [%rd97], %r182;
|
|
ld.global.u32 %r177, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r177, %f12, %f13, %f14, %f63, %f64, %f65, %r399, %f622, %f623, %rd90, %r110);
|
|
// inline asm
|
|
add.s32 %r183, %r398, 2;
|
|
cvt.s64.s32 %rd93, %r183;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd307, %r399, %r125, %rd93, %rd306, %rd306, %rd306);
|
|
// inline asm
|
|
ld.f32 %f631, [%rd91];
|
|
ld.local.f32 %f632, [%rd97];
|
|
mul.f32 %f633, %f632, %f631;
|
|
ld.f32 %f634, [%rd91+4];
|
|
mul.f32 %f635, %f632, %f634;
|
|
ld.f32 %f636, [%rd91+8];
|
|
mul.f32 %f637, %f632, %f636;
|
|
mul.f32 %f638, %f82, 0f3EA2F983;
|
|
mul.f32 %f639, %f638, %f633;
|
|
mul.f32 %f640, %f638, %f635;
|
|
mul.f32 %f641, %f638, %f637;
|
|
fma.rn.f32 %f52, %f624, %f639, %f52;
|
|
fma.rn.f32 %f51, %f624, %f640, %f51;
|
|
fma.rn.f32 %f50, %f624, %f641, %f50;
|
|
mul.f32 %f642, %f24, %f64;
|
|
fma.rn.f32 %f643, %f22, %f63, %f642;
|
|
fma.rn.f32 %f644, %f26, %f65, %f643;
|
|
cvt.sat.f32.f32 %f645, %f644;
|
|
fma.rn.f32 %f55, %f639, %f645, %f55;
|
|
fma.rn.f32 %f54, %f640, %f645, %f54;
|
|
fma.rn.f32 %f53, %f645, %f641, %f53;
|
|
mul.f32 %f646, %f34, %f64;
|
|
fma.rn.f32 %f647, %f33, %f63, %f646;
|
|
fma.rn.f32 %f648, %f35, %f65, %f647;
|
|
cvt.sat.f32.f32 %f649, %f648;
|
|
fma.rn.f32 %f58, %f639, %f649, %f58;
|
|
fma.rn.f32 %f57, %f640, %f649, %f57;
|
|
fma.rn.f32 %f56, %f641, %f649, %f56;
|
|
mul.f32 %f650, %f43, %f64;
|
|
fma.rn.f32 %f651, %f42, %f63, %f650;
|
|
fma.rn.f32 %f652, %f44, %f65, %f651;
|
|
cvt.sat.f32.f32 %f653, %f652;
|
|
fma.rn.f32 %f61, %f639, %f653, %f61;
|
|
fma.rn.f32 %f60, %f640, %f653, %f60;
|
|
fma.rn.f32 %f59, %f641, %f653, %f59;
|
|
add.f32 %f49, %f49, %f632;
|
|
|
|
BB0_28:
|
|
add.s32 %r15, %r15, 1;
|
|
|
|
BB0_29:
|
|
add.s32 %r400, %r400, 1;
|
|
setp.lt.s32 %p43, %r400, %r2;
|
|
@%p43 bra BB0_12;
|
|
bra.uni BB0_30;
|
|
|
|
BB0_10:
|
|
mov.f32 %f50, %f49;
|
|
mov.f32 %f51, %f49;
|
|
mov.f32 %f52, %f49;
|
|
mov.f32 %f53, %f49;
|
|
mov.f32 %f54, %f49;
|
|
mov.f32 %f55, %f49;
|
|
mov.f32 %f56, %f49;
|
|
mov.f32 %f57, %f49;
|
|
mov.f32 %f58, %f49;
|
|
mov.f32 %f59, %f49;
|
|
mov.f32 %f60, %f49;
|
|
mov.f32 %f61, %f49;
|
|
|
|
BB0_30:
|
|
mov.pred %p156, %p18;
|
|
|
|
BB0_31:
|
|
cvt.rn.f32.s32 %f654, %r15;
|
|
mov.f32 %f655, 0f3F800000;
|
|
max.f32 %f656, %f654, %f655;
|
|
rcp.rn.f32 %f657, %f656;
|
|
mul.f32 %f1368, %f52, %f657;
|
|
mul.f32 %f1369, %f51, %f657;
|
|
mul.f32 %f1370, %f50, %f657;
|
|
div.rn.f32 %f1371, %f49, %f656;
|
|
mul.f32 %f1365, %f55, %f657;
|
|
mul.f32 %f1366, %f54, %f657;
|
|
mul.f32 %f1367, %f53, %f657;
|
|
mul.f32 %f1362, %f58, %f657;
|
|
mul.f32 %f1363, %f57, %f657;
|
|
mul.f32 %f1364, %f56, %f657;
|
|
mul.f32 %f1359, %f61, %f657;
|
|
mul.f32 %f1360, %f60, %f657;
|
|
mul.f32 %f1361, %f59, %f657;
|
|
@!%p156 bra BB0_84;
|
|
bra.uni BB0_32;
|
|
|
|
BB0_32:
|
|
abs.f32 %f1242, %f1246;
|
|
abs.f32 %f1241, %f1244;
|
|
setp.gt.f32 %p44, %f1241, %f1242;
|
|
neg.f32 %f670, %f1245;
|
|
selp.f32 %f671, %f670, 0f00000000, %p44;
|
|
neg.f32 %f672, %f1246;
|
|
selp.f32 %f673, %f1244, %f672, %p44;
|
|
selp.f32 %f674, 0f00000000, %f1245, %p44;
|
|
mul.f32 %f675, %f673, %f673;
|
|
fma.rn.f32 %f676, %f671, %f671, %f675;
|
|
fma.rn.f32 %f677, %f674, %f674, %f676;
|
|
sqrt.rn.f32 %f678, %f677;
|
|
rcp.rn.f32 %f679, %f678;
|
|
mul.f32 %f148, %f671, %f679;
|
|
mul.f32 %f149, %f673, %f679;
|
|
mul.f32 %f150, %f674, %f679;
|
|
mov.f32 %f669, 0f00000000;
|
|
setp.lt.s32 %p45, %r3, 1;
|
|
mov.f32 %f1323, %f669;
|
|
mov.f32 %f1324, %f669;
|
|
mov.f32 %f1325, %f669;
|
|
mov.f32 %f1326, %f669;
|
|
mov.f32 %f1327, %f669;
|
|
mov.f32 %f1328, %f669;
|
|
mov.f32 %f1329, %f669;
|
|
mov.f32 %f1330, %f669;
|
|
mov.f32 %f1331, %f669;
|
|
mov.f32 %f1332, %f669;
|
|
mov.f32 %f1333, %f669;
|
|
mov.f32 %f1334, %f669;
|
|
@%p45 bra BB0_83;
|
|
|
|
cvt.rn.f32.s32 %f692, %r3;
|
|
rcp.rn.f32 %f151, %f692;
|
|
add.u64 %rd99, %SP, 16;
|
|
cvta.to.local.u64 %rd6, %rd99;
|
|
mul.f32 %f152, %f12, 0f3456BF95;
|
|
mul.f32 %f153, %f13, 0f3456BF95;
|
|
mul.f32 %f154, %f14, 0f3456BF95;
|
|
add.u64 %rd100, %SP, 0;
|
|
cvta.to.local.u64 %rd7, %rd100;
|
|
mul.f32 %f693, %f1244, %f149;
|
|
mul.f32 %f694, %f1245, %f148;
|
|
sub.f32 %f155, %f694, %f693;
|
|
mul.f32 %f695, %f1246, %f148;
|
|
mul.f32 %f696, %f1244, %f150;
|
|
sub.f32 %f156, %f696, %f695;
|
|
mul.f32 %f697, %f1245, %f150;
|
|
mul.f32 %f698, %f1246, %f149;
|
|
sub.f32 %f157, %f698, %f697;
|
|
mov.f32 %f1323, 0f00000000;
|
|
mov.u32 %r184, 0;
|
|
abs.f32 %f699, %f153;
|
|
abs.f32 %f700, %f152;
|
|
max.f32 %f701, %f700, %f699;
|
|
abs.f32 %f702, %f154;
|
|
max.f32 %f703, %f701, %f702;
|
|
mov.u32 %r404, %r184;
|
|
mov.f32 %f1324, %f1323;
|
|
mov.f32 %f1325, %f1323;
|
|
mov.f32 %f1326, %f1323;
|
|
mov.f32 %f1327, %f1323;
|
|
mov.f32 %f1328, %f1323;
|
|
mov.f32 %f1329, %f1323;
|
|
mov.f32 %f1330, %f1323;
|
|
mov.f32 %f1331, %f1323;
|
|
mov.f32 %f1332, %f1323;
|
|
mov.f32 %f1333, %f1323;
|
|
mov.f32 %f1334, %f1323;
|
|
|
|
BB0_34:
|
|
cvt.rn.f32.s32 %f170, %r404;
|
|
max.f32 %f171, %f703, %f367;
|
|
mov.u32 %r406, %r184;
|
|
|
|
BB0_35:
|
|
mad.lo.s32 %r186, %r407, 1664525, 1013904223;
|
|
and.b32 %r187, %r186, 16777215;
|
|
cvt.rn.f32.u32 %f705, %r187;
|
|
fma.rn.f32 %f706, %f705, 0f33800000, %f170;
|
|
mul.f32 %f184, %f151, %f706;
|
|
mad.lo.s32 %r407, %r186, 1664525, 1013904223;
|
|
and.b32 %r188, %r407, 16777215;
|
|
cvt.rn.f32.u32 %f707, %r188;
|
|
cvt.rn.f32.s32 %f708, %r406;
|
|
fma.rn.f32 %f709, %f707, 0f33800000, %f708;
|
|
mul.f32 %f710, %f151, %f709;
|
|
mul.f32 %f711, %f184, %f184;
|
|
sub.f32 %f713, %f655, %f711;
|
|
mov.f32 %f714, 0f00000000;
|
|
max.f32 %f715, %f714, %f713;
|
|
sqrt.rn.f32 %f185, %f715;
|
|
mul.f32 %f1341, %f710, 0f40C90FDB;
|
|
abs.f32 %f187, %f1341;
|
|
setp.neu.f32 %p46, %f187, 0f7F800000;
|
|
mov.f32 %f1335, %f1341;
|
|
@%p46 bra BB0_37;
|
|
|
|
mul.rn.f32 %f1335, %f1341, %f714;
|
|
|
|
BB0_37:
|
|
mul.f32 %f717, %f1335, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r417, %f717;
|
|
cvt.rn.f32.s32 %f718, %r417;
|
|
neg.f32 %f719, %f718;
|
|
mov.f32 %f720, 0f3FC90FDA;
|
|
fma.rn.f32 %f721, %f719, %f720, %f1335;
|
|
mov.f32 %f722, 0f33A22168;
|
|
fma.rn.f32 %f723, %f719, %f722, %f721;
|
|
mov.f32 %f724, 0f27C234C5;
|
|
fma.rn.f32 %f1336, %f719, %f724, %f723;
|
|
abs.f32 %f725, %f1335;
|
|
setp.leu.f32 %p47, %f725, 0f47CE4780;
|
|
@%p47 bra BB0_48;
|
|
|
|
mov.b32 %r27, %f1335;
|
|
shr.u32 %r28, %r27, 23;
|
|
shl.b32 %r191, %r27, 8;
|
|
or.b32 %r29, %r191, -2147483648;
|
|
mov.u32 %r408, 0;
|
|
mov.u64 %rd310, 0;
|
|
mov.u64 %rd309, %rd6;
|
|
mov.u32 %r409, %r408;
|
|
|
|
BB0_39:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd102, %rd310, 2;
|
|
mov.u64 %rd103, __cudart_i2opi_f;
|
|
add.s64 %rd104, %rd103, %rd102;
|
|
ld.const.u32 %r194, [%rd104];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r192, %r194, %r29, %r409;
|
|
madc.hi.u32 %r409, %r194, %r29, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd309], %r192;
|
|
add.s32 %r408, %r408, 1;
|
|
cvt.s64.s32 %rd310, %r408;
|
|
mul.wide.s32 %rd105, %r408, 4;
|
|
add.s64 %rd309, %rd6, %rd105;
|
|
setp.ne.s32 %p48, %r408, 6;
|
|
@%p48 bra BB0_39;
|
|
|
|
and.b32 %r197, %r28, 255;
|
|
add.s32 %r198, %r197, -128;
|
|
shr.u32 %r199, %r198, 5;
|
|
and.b32 %r34, %r27, -2147483648;
|
|
st.local.u32 [%rd6+24], %r409;
|
|
mov.u32 %r200, 6;
|
|
sub.s32 %r201, %r200, %r199;
|
|
mul.wide.s32 %rd106, %r201, 4;
|
|
add.s64 %rd13, %rd6, %rd106;
|
|
ld.local.u32 %r410, [%rd13];
|
|
ld.local.u32 %r411, [%rd13+-4];
|
|
and.b32 %r37, %r28, 31;
|
|
setp.eq.s32 %p49, %r37, 0;
|
|
@%p49 bra BB0_42;
|
|
|
|
mov.u32 %r202, 32;
|
|
sub.s32 %r203, %r202, %r37;
|
|
shr.u32 %r204, %r411, %r203;
|
|
shl.b32 %r205, %r410, %r37;
|
|
add.s32 %r410, %r204, %r205;
|
|
ld.local.u32 %r206, [%rd13+-8];
|
|
shr.u32 %r207, %r206, %r203;
|
|
shl.b32 %r208, %r411, %r37;
|
|
add.s32 %r411, %r207, %r208;
|
|
|
|
BB0_42:
|
|
shr.u32 %r209, %r411, 30;
|
|
shl.b32 %r210, %r410, 2;
|
|
add.s32 %r412, %r209, %r210;
|
|
shl.b32 %r43, %r411, 2;
|
|
shr.u32 %r211, %r412, 31;
|
|
shr.u32 %r212, %r410, 30;
|
|
add.s32 %r44, %r211, %r212;
|
|
setp.eq.s32 %p50, %r211, 0;
|
|
@%p50 bra BB0_43;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_43:
|
|
mov.u32 %r413, %r34;
|
|
mov.u32 %r414, %r43;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_44:
|
|
not.b32 %r213, %r412;
|
|
neg.s32 %r414, %r43;
|
|
setp.eq.s32 %p51, %r43, 0;
|
|
selp.u32 %r214, 1, 0, %p51;
|
|
add.s32 %r412, %r214, %r213;
|
|
xor.b32 %r413, %r34, -2147483648;
|
|
|
|
BB0_45:
|
|
clz.b32 %r416, %r412;
|
|
setp.eq.s32 %p52, %r416, 0;
|
|
shl.b32 %r215, %r412, %r416;
|
|
mov.u32 %r216, 32;
|
|
sub.s32 %r217, %r216, %r416;
|
|
shr.u32 %r218, %r414, %r217;
|
|
add.s32 %r219, %r218, %r215;
|
|
selp.b32 %r52, %r412, %r219, %p52;
|
|
mov.u32 %r220, -921707870;
|
|
mul.hi.u32 %r415, %r52, %r220;
|
|
setp.eq.s32 %p53, %r34, 0;
|
|
neg.s32 %r221, %r44;
|
|
selp.b32 %r417, %r44, %r221, %p53;
|
|
setp.lt.s32 %p54, %r415, 1;
|
|
@%p54 bra BB0_47;
|
|
|
|
mul.lo.s32 %r222, %r52, -921707870;
|
|
shr.u32 %r223, %r222, 31;
|
|
shl.b32 %r224, %r415, 1;
|
|
add.s32 %r415, %r223, %r224;
|
|
add.s32 %r416, %r416, 1;
|
|
|
|
BB0_47:
|
|
mov.u32 %r225, 126;
|
|
sub.s32 %r226, %r225, %r416;
|
|
shl.b32 %r227, %r226, 23;
|
|
add.s32 %r228, %r415, 1;
|
|
shr.u32 %r229, %r228, 7;
|
|
add.s32 %r230, %r229, 1;
|
|
shr.u32 %r231, %r230, 1;
|
|
add.s32 %r232, %r231, %r227;
|
|
or.b32 %r233, %r232, %r413;
|
|
mov.b32 %f1336, %r233;
|
|
|
|
BB0_48:
|
|
mul.rn.f32 %f193, %f1336, %f1336;
|
|
add.s32 %r60, %r417, 1;
|
|
and.b32 %r61, %r60, 1;
|
|
setp.eq.s32 %p55, %r61, 0;
|
|
@%p55 bra BB0_50;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_50:
|
|
mov.f32 %f728, 0f3C08839E;
|
|
mov.f32 %f729, 0fB94CA1F9;
|
|
fma.rn.f32 %f1337, %f729, %f193, %f728;
|
|
bra.uni BB0_51;
|
|
|
|
BB0_49:
|
|
mov.f32 %f726, 0fBAB6061A;
|
|
mov.f32 %f727, 0f37CCF5CE;
|
|
fma.rn.f32 %f1337, %f727, %f193, %f726;
|
|
|
|
BB0_51:
|
|
@%p55 bra BB0_53;
|
|
bra.uni BB0_52;
|
|
|
|
BB0_53:
|
|
mov.f32 %f733, 0fBE2AAAA3;
|
|
fma.rn.f32 %f734, %f1337, %f193, %f733;
|
|
fma.rn.f32 %f1338, %f734, %f193, %f714;
|
|
bra.uni BB0_54;
|
|
|
|
BB0_52:
|
|
mov.f32 %f730, 0f3D2AAAA5;
|
|
fma.rn.f32 %f731, %f1337, %f193, %f730;
|
|
mov.f32 %f732, 0fBF000000;
|
|
fma.rn.f32 %f1338, %f731, %f193, %f732;
|
|
|
|
BB0_54:
|
|
fma.rn.f32 %f1339, %f1338, %f1336, %f1336;
|
|
@%p55 bra BB0_56;
|
|
|
|
fma.rn.f32 %f1339, %f1338, %f193, %f655;
|
|
|
|
BB0_56:
|
|
and.b32 %r234, %r60, 2;
|
|
setp.eq.s32 %p58, %r234, 0;
|
|
@%p58 bra BB0_58;
|
|
|
|
mov.f32 %f738, 0fBF800000;
|
|
fma.rn.f32 %f1339, %f1339, %f738, %f714;
|
|
|
|
BB0_58:
|
|
@%p46 bra BB0_60;
|
|
|
|
mul.rn.f32 %f1341, %f1341, %f714;
|
|
|
|
BB0_60:
|
|
mul.f32 %f740, %f1341, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r427, %f740;
|
|
cvt.rn.f32.s32 %f741, %r427;
|
|
neg.f32 %f742, %f741;
|
|
fma.rn.f32 %f744, %f742, %f720, %f1341;
|
|
fma.rn.f32 %f746, %f742, %f722, %f744;
|
|
fma.rn.f32 %f1342, %f742, %f724, %f746;
|
|
abs.f32 %f748, %f1341;
|
|
setp.leu.f32 %p60, %f748, 0f47CE4780;
|
|
@%p60 bra BB0_71;
|
|
|
|
mov.b32 %r63, %f1341;
|
|
shr.u32 %r64, %r63, 23;
|
|
shl.b32 %r237, %r63, 8;
|
|
or.b32 %r65, %r237, -2147483648;
|
|
mov.u32 %r418, 0;
|
|
mov.u64 %rd311, %rd6;
|
|
mov.u64 %rd312, %rd30;
|
|
mov.u32 %r419, %r418;
|
|
|
|
BB0_62:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd108, %rd312, 2;
|
|
mov.u64 %rd109, __cudart_i2opi_f;
|
|
add.s64 %rd110, %rd109, %rd108;
|
|
ld.const.u32 %r240, [%rd110];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r238, %r240, %r65, %r419;
|
|
madc.hi.u32 %r419, %r240, %r65, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd311], %r238;
|
|
add.s32 %r418, %r418, 1;
|
|
cvt.s64.s32 %rd312, %r418;
|
|
mul.wide.s32 %rd111, %r418, 4;
|
|
add.s64 %rd311, %rd6, %rd111;
|
|
setp.ne.s32 %p61, %r418, 6;
|
|
@%p61 bra BB0_62;
|
|
|
|
and.b32 %r243, %r64, 255;
|
|
add.s32 %r244, %r243, -128;
|
|
shr.u32 %r245, %r244, 5;
|
|
and.b32 %r70, %r63, -2147483648;
|
|
st.local.u32 [%rd6+24], %r419;
|
|
mov.u32 %r246, 6;
|
|
sub.s32 %r247, %r246, %r245;
|
|
mul.wide.s32 %rd112, %r247, 4;
|
|
add.s64 %rd19, %rd6, %rd112;
|
|
ld.local.u32 %r420, [%rd19];
|
|
ld.local.u32 %r421, [%rd19+-4];
|
|
and.b32 %r73, %r64, 31;
|
|
setp.eq.s32 %p62, %r73, 0;
|
|
@%p62 bra BB0_65;
|
|
|
|
mov.u32 %r248, 32;
|
|
sub.s32 %r249, %r248, %r73;
|
|
shr.u32 %r250, %r421, %r249;
|
|
shl.b32 %r251, %r420, %r73;
|
|
add.s32 %r420, %r250, %r251;
|
|
ld.local.u32 %r252, [%rd19+-8];
|
|
shr.u32 %r253, %r252, %r249;
|
|
shl.b32 %r254, %r421, %r73;
|
|
add.s32 %r421, %r253, %r254;
|
|
|
|
BB0_65:
|
|
shr.u32 %r255, %r421, 30;
|
|
shl.b32 %r256, %r420, 2;
|
|
add.s32 %r422, %r255, %r256;
|
|
shl.b32 %r79, %r421, 2;
|
|
shr.u32 %r257, %r422, 31;
|
|
shr.u32 %r258, %r420, 30;
|
|
add.s32 %r80, %r257, %r258;
|
|
setp.eq.s32 %p63, %r257, 0;
|
|
@%p63 bra BB0_66;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_66:
|
|
mov.u32 %r423, %r70;
|
|
mov.u32 %r424, %r79;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_67:
|
|
not.b32 %r259, %r422;
|
|
neg.s32 %r424, %r79;
|
|
setp.eq.s32 %p64, %r79, 0;
|
|
selp.u32 %r260, 1, 0, %p64;
|
|
add.s32 %r422, %r260, %r259;
|
|
xor.b32 %r423, %r70, -2147483648;
|
|
|
|
BB0_68:
|
|
clz.b32 %r426, %r422;
|
|
setp.eq.s32 %p65, %r426, 0;
|
|
shl.b32 %r261, %r422, %r426;
|
|
mov.u32 %r262, 32;
|
|
sub.s32 %r263, %r262, %r426;
|
|
shr.u32 %r264, %r424, %r263;
|
|
add.s32 %r265, %r264, %r261;
|
|
selp.b32 %r88, %r422, %r265, %p65;
|
|
mov.u32 %r266, -921707870;
|
|
mul.hi.u32 %r425, %r88, %r266;
|
|
setp.eq.s32 %p66, %r70, 0;
|
|
neg.s32 %r267, %r80;
|
|
selp.b32 %r427, %r80, %r267, %p66;
|
|
setp.lt.s32 %p67, %r425, 1;
|
|
@%p67 bra BB0_70;
|
|
|
|
mul.lo.s32 %r268, %r88, -921707870;
|
|
shr.u32 %r269, %r268, 31;
|
|
shl.b32 %r270, %r425, 1;
|
|
add.s32 %r425, %r269, %r270;
|
|
add.s32 %r426, %r426, 1;
|
|
|
|
BB0_70:
|
|
mov.u32 %r271, 126;
|
|
sub.s32 %r272, %r271, %r426;
|
|
shl.b32 %r273, %r272, 23;
|
|
add.s32 %r274, %r425, 1;
|
|
shr.u32 %r275, %r274, 7;
|
|
add.s32 %r276, %r275, 1;
|
|
shr.u32 %r277, %r276, 1;
|
|
add.s32 %r278, %r277, %r273;
|
|
or.b32 %r279, %r278, %r423;
|
|
mov.b32 %f1342, %r279;
|
|
|
|
BB0_71:
|
|
mul.rn.f32 %f210, %f1342, %f1342;
|
|
and.b32 %r96, %r427, 1;
|
|
setp.eq.s32 %p68, %r96, 0;
|
|
@%p68 bra BB0_73;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_73:
|
|
mov.f32 %f751, 0f3C08839E;
|
|
mov.f32 %f752, 0fB94CA1F9;
|
|
fma.rn.f32 %f1343, %f752, %f210, %f751;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_72:
|
|
mov.f32 %f749, 0fBAB6061A;
|
|
mov.f32 %f750, 0f37CCF5CE;
|
|
fma.rn.f32 %f1343, %f750, %f210, %f749;
|
|
|
|
BB0_74:
|
|
@%p68 bra BB0_76;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_76:
|
|
mov.f32 %f756, 0fBE2AAAA3;
|
|
fma.rn.f32 %f757, %f1343, %f210, %f756;
|
|
fma.rn.f32 %f1344, %f757, %f210, %f714;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_75:
|
|
mov.f32 %f753, 0f3D2AAAA5;
|
|
fma.rn.f32 %f754, %f1343, %f210, %f753;
|
|
mov.f32 %f755, 0fBF000000;
|
|
fma.rn.f32 %f1344, %f754, %f210, %f755;
|
|
|
|
BB0_77:
|
|
fma.rn.f32 %f1345, %f1344, %f1342, %f1342;
|
|
@%p68 bra BB0_79;
|
|
|
|
fma.rn.f32 %f1345, %f1344, %f210, %f655;
|
|
|
|
BB0_79:
|
|
and.b32 %r280, %r427, 2;
|
|
setp.eq.s32 %p71, %r280, 0;
|
|
@%p71 bra BB0_81;
|
|
|
|
mov.f32 %f761, 0fBF800000;
|
|
fma.rn.f32 %f1345, %f1345, %f761, %f714;
|
|
|
|
BB0_81:
|
|
mul.f32 %f770, %f185, %f1339;
|
|
mul.f32 %f771, %f185, %f1345;
|
|
mul.f32 %f772, %f148, %f771;
|
|
mul.f32 %f773, %f149, %f771;
|
|
mul.f32 %f774, %f150, %f771;
|
|
fma.rn.f32 %f775, %f157, %f770, %f772;
|
|
fma.rn.f32 %f776, %f156, %f770, %f773;
|
|
fma.rn.f32 %f777, %f155, %f770, %f774;
|
|
fma.rn.f32 %f765, %f1244, %f184, %f775;
|
|
fma.rn.f32 %f766, %f1245, %f184, %f776;
|
|
fma.rn.f32 %f767, %f1246, %f184, %f777;
|
|
mov.u32 %r282, 0;
|
|
st.local.u32 [%rd7+8], %r282;
|
|
st.local.u32 [%rd7+4], %r282;
|
|
st.local.u32 [%rd7], %r282;
|
|
ld.global.u32 %r281, [root];
|
|
mov.f32 %f769, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r281, %f12, %f13, %f14, %f765, %f766, %f767, %r282, %f171, %f769, %rd100, %r125);
|
|
// inline asm
|
|
ld.local.f32 %f778, [%rd7];
|
|
max.f32 %f780, %f778, %f714;
|
|
ld.local.f32 %f781, [%rd7+4];
|
|
max.f32 %f782, %f781, %f714;
|
|
ld.local.f32 %f783, [%rd7+8];
|
|
max.f32 %f784, %f783, %f714;
|
|
mul.f32 %f785, %f24, %f766;
|
|
fma.rn.f32 %f786, %f22, %f765, %f785;
|
|
fma.rn.f32 %f787, %f26, %f767, %f786;
|
|
cvt.sat.f32.f32 %f788, %f787;
|
|
fma.rn.f32 %f1328, %f780, %f788, %f1328;
|
|
fma.rn.f32 %f1327, %f782, %f788, %f1327;
|
|
fma.rn.f32 %f1326, %f784, %f788, %f1326;
|
|
mul.f32 %f789, %f34, %f766;
|
|
fma.rn.f32 %f790, %f33, %f765, %f789;
|
|
fma.rn.f32 %f791, %f35, %f767, %f790;
|
|
cvt.sat.f32.f32 %f792, %f791;
|
|
fma.rn.f32 %f1331, %f780, %f792, %f1331;
|
|
fma.rn.f32 %f1330, %f782, %f792, %f1330;
|
|
fma.rn.f32 %f1329, %f784, %f792, %f1329;
|
|
mul.f32 %f793, %f43, %f766;
|
|
fma.rn.f32 %f794, %f42, %f765, %f793;
|
|
fma.rn.f32 %f795, %f44, %f767, %f794;
|
|
cvt.sat.f32.f32 %f796, %f795;
|
|
fma.rn.f32 %f1334, %f780, %f796, %f1334;
|
|
fma.rn.f32 %f1333, %f782, %f796, %f1333;
|
|
fma.rn.f32 %f1332, %f784, %f796, %f1332;
|
|
mul.f32 %f797, %f1245, %f766;
|
|
fma.rn.f32 %f798, %f1244, %f765, %f797;
|
|
fma.rn.f32 %f799, %f1246, %f767, %f798;
|
|
cvt.sat.f32.f32 %f800, %f799;
|
|
fma.rn.f32 %f1325, %f780, %f800, %f1325;
|
|
fma.rn.f32 %f1324, %f782, %f800, %f1324;
|
|
fma.rn.f32 %f1323, %f784, %f800, %f1323;
|
|
add.s32 %r406, %r406, 1;
|
|
setp.lt.s32 %p72, %r406, %r3;
|
|
@%p72 bra BB0_35;
|
|
|
|
add.s32 %r404, %r404, 1;
|
|
setp.lt.s32 %p73, %r404, %r3;
|
|
@%p73 bra BB0_34;
|
|
|
|
BB0_83:
|
|
mul.lo.s32 %r284, %r3, %r3;
|
|
cvt.rn.f32.s32 %f801, %r284;
|
|
rcp.rn.f32 %f802, %f801;
|
|
mul.f32 %f803, %f1325, %f802;
|
|
mul.f32 %f804, %f1324, %f802;
|
|
mul.f32 %f805, %f1323, %f802;
|
|
div.rn.f32 %f1371, %f669, %f801;
|
|
mul.f32 %f1365, %f1328, %f802;
|
|
mul.f32 %f1366, %f1327, %f802;
|
|
mul.f32 %f1367, %f1326, %f802;
|
|
mul.f32 %f1362, %f1331, %f802;
|
|
mul.f32 %f1363, %f1330, %f802;
|
|
mul.f32 %f1364, %f1329, %f802;
|
|
mul.f32 %f1359, %f1334, %f802;
|
|
mul.f32 %f1360, %f1333, %f802;
|
|
mul.f32 %f1361, %f1332, %f802;
|
|
fma.rn.f32 %f1368, %f1325, %f802, %f803;
|
|
fma.rn.f32 %f1369, %f1324, %f802, %f804;
|
|
fma.rn.f32 %f1370, %f1323, %f802, %f805;
|
|
|
|
BB0_84:
|
|
setp.gt.s32 %p74, %r1, -1;
|
|
selp.f32 %f272, %f1371, %f1368, %p74;
|
|
ld.global.u32 %r429, [imageEnabled];
|
|
and.b32 %r285, %r429, 8;
|
|
setp.eq.s32 %p75, %r285, 0;
|
|
@%p75 bra BB0_97;
|
|
|
|
cvt.u64.u32 %rd116, %r5;
|
|
cvt.u64.u32 %rd117, %r6;
|
|
mov.u64 %rd120, image_Mask;
|
|
cvta.global.u64 %rd115, %rd120;
|
|
// inline asm
|
|
call (%rd114), _rt_buffer_get_64, (%rd115, %r109, %r109, %rd116, %rd117, %rd30, %rd30);
|
|
// inline asm
|
|
mov.f32 %f809, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f810, %f809;
|
|
fma.rn.f32 %f811, %f810, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f273, %f811;
|
|
abs.f32 %f274, %f272;
|
|
setp.lt.f32 %p76, %f274, 0f00800000;
|
|
mul.f32 %f812, %f274, 0f4B800000;
|
|
selp.f32 %f813, 0fC3170000, 0fC2FE0000, %p76;
|
|
selp.f32 %f814, %f812, %f274, %p76;
|
|
mov.b32 %r288, %f814;
|
|
and.b32 %r289, %r288, 8388607;
|
|
or.b32 %r290, %r289, 1065353216;
|
|
mov.b32 %f815, %r290;
|
|
shr.u32 %r291, %r288, 23;
|
|
cvt.rn.f32.u32 %f816, %r291;
|
|
add.f32 %f817, %f813, %f816;
|
|
setp.gt.f32 %p77, %f815, 0f3FB504F3;
|
|
mul.f32 %f818, %f815, 0f3F000000;
|
|
add.f32 %f819, %f817, 0f3F800000;
|
|
selp.f32 %f820, %f818, %f815, %p77;
|
|
selp.f32 %f821, %f819, %f817, %p77;
|
|
add.f32 %f822, %f820, 0fBF800000;
|
|
add.f32 %f808, %f820, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f807,%f808;
|
|
// inline asm
|
|
add.f32 %f823, %f822, %f822;
|
|
mul.f32 %f824, %f807, %f823;
|
|
mul.f32 %f825, %f824, %f824;
|
|
mov.f32 %f826, 0f3C4CAF63;
|
|
mov.f32 %f827, 0f3B18F0FE;
|
|
fma.rn.f32 %f828, %f827, %f825, %f826;
|
|
mov.f32 %f829, 0f3DAAAABD;
|
|
fma.rn.f32 %f830, %f828, %f825, %f829;
|
|
mul.rn.f32 %f831, %f830, %f825;
|
|
mul.rn.f32 %f832, %f831, %f824;
|
|
sub.f32 %f833, %f822, %f824;
|
|
neg.f32 %f834, %f824;
|
|
add.f32 %f835, %f833, %f833;
|
|
fma.rn.f32 %f836, %f834, %f822, %f835;
|
|
mul.rn.f32 %f837, %f807, %f836;
|
|
add.f32 %f838, %f832, %f824;
|
|
sub.f32 %f839, %f824, %f838;
|
|
add.f32 %f840, %f832, %f839;
|
|
add.f32 %f841, %f837, %f840;
|
|
add.f32 %f842, %f838, %f841;
|
|
sub.f32 %f843, %f838, %f842;
|
|
add.f32 %f844, %f841, %f843;
|
|
mov.f32 %f845, 0f3F317200;
|
|
mul.rn.f32 %f846, %f821, %f845;
|
|
mov.f32 %f847, 0f35BFBE8E;
|
|
mul.rn.f32 %f848, %f821, %f847;
|
|
add.f32 %f849, %f846, %f842;
|
|
sub.f32 %f850, %f846, %f849;
|
|
add.f32 %f851, %f842, %f850;
|
|
add.f32 %f852, %f844, %f851;
|
|
add.f32 %f853, %f848, %f852;
|
|
add.f32 %f854, %f849, %f853;
|
|
sub.f32 %f855, %f849, %f854;
|
|
add.f32 %f856, %f853, %f855;
|
|
mov.f32 %f857, 0f3EE8BA2E;
|
|
mul.rn.f32 %f858, %f857, %f854;
|
|
neg.f32 %f859, %f858;
|
|
fma.rn.f32 %f860, %f857, %f854, %f859;
|
|
fma.rn.f32 %f861, %f857, %f856, %f860;
|
|
mov.f32 %f862, 0f00000000;
|
|
fma.rn.f32 %f863, %f862, %f854, %f861;
|
|
add.rn.f32 %f864, %f858, %f863;
|
|
neg.f32 %f865, %f864;
|
|
add.rn.f32 %f866, %f858, %f865;
|
|
add.rn.f32 %f867, %f866, %f863;
|
|
mov.b32 %r292, %f864;
|
|
setp.eq.s32 %p78, %r292, 1118925336;
|
|
add.s32 %r293, %r292, -1;
|
|
mov.b32 %f868, %r293;
|
|
add.f32 %f869, %f867, 0f37000000;
|
|
selp.f32 %f870, %f868, %f864, %p78;
|
|
selp.f32 %f275, %f869, %f867, %p78;
|
|
mul.f32 %f871, %f870, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f872, %f871;
|
|
mov.f32 %f873, 0fBF317200;
|
|
fma.rn.f32 %f874, %f872, %f873, %f870;
|
|
mov.f32 %f875, 0fB5BFBE8E;
|
|
fma.rn.f32 %f876, %f872, %f875, %f874;
|
|
mul.f32 %f877, %f876, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f878, %f877;
|
|
add.f32 %f879, %f872, 0f00000000;
|
|
ex2.approx.f32 %f880, %f879;
|
|
mul.f32 %f881, %f878, %f880;
|
|
setp.lt.f32 %p79, %f870, 0fC2D20000;
|
|
selp.f32 %f882, 0f00000000, %f881, %p79;
|
|
setp.gt.f32 %p80, %f870, 0f42D20000;
|
|
selp.f32 %f1372, 0f7F800000, %f882, %p80;
|
|
setp.eq.f32 %p81, %f1372, 0f7F800000;
|
|
@%p81 bra BB0_87;
|
|
|
|
fma.rn.f32 %f1372, %f1372, %f275, %f1372;
|
|
|
|
BB0_87:
|
|
setp.lt.f32 %p82, %f272, 0f00000000;
|
|
setp.eq.f32 %p83, %f273, 0f3F800000;
|
|
and.pred %p3, %p82, %p83;
|
|
mov.b32 %r294, %f1372;
|
|
xor.b32 %r295, %r294, -2147483648;
|
|
mov.b32 %f883, %r295;
|
|
selp.f32 %f1374, %f883, %f1372, %p3;
|
|
setp.eq.f32 %p84, %f272, 0f00000000;
|
|
@%p84 bra BB0_90;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_90:
|
|
add.f32 %f886, %f272, %f272;
|
|
selp.f32 %f1374, %f886, 0f00000000, %p83;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd241, image_HDR;
|
|
cvta.global.u64 %rd236, %rd241;
|
|
mov.u32 %r374, 8;
|
|
mov.u64 %rd240, 0;
|
|
// inline asm
|
|
call (%rd235), _rt_buffer_get_64, (%rd236, %r109, %r374, %rd23, %rd24, %rd240, %rd240);
|
|
// inline asm
|
|
mov.f32 %f1213, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f1213;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs118, 0;
|
|
st.v4.u16 [%rd235], {%rs117, %rs117, %rs117, %rs118};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r375, [additive];
|
|
setp.eq.s32 %p153, %r375, 0;
|
|
@%p153 bra BB0_155;
|
|
|
|
mov.u64 %rd254, image_RNM0;
|
|
cvta.global.u64 %rd243, %rd254;
|
|
mov.u32 %r379, 8;
|
|
mov.u64 %rd253, 0;
|
|
// inline asm
|
|
call (%rd242), _rt_buffer_get_64, (%rd243, %r109, %r379, %rd23, %rd24, %rd253, %rd253);
|
|
// inline asm
|
|
ld.v4.u16 {%rs125, %rs126, %rs127, %rs128}, [%rd242];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1214, %rs125;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1215, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1216, %rs127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd248), _rt_buffer_get_64, (%rd243, %r109, %r379, %rd23, %rd24, %rd253, %rd253);
|
|
// inline asm
|
|
add.f32 %f1217, %f1214, 0f00000000;
|
|
add.f32 %f1218, %f1215, 0f00000000;
|
|
add.f32 %f1219, %f1216, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1219;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1218;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f1217;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs129, 0;
|
|
st.v4.u16 [%rd248], {%rs122, %rs123, %rs124, %rs129};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd261, image_RNM0;
|
|
cvta.global.u64 %rd256, %rd261;
|
|
mov.u32 %r381, 8;
|
|
mov.u64 %rd260, 0;
|
|
// inline asm
|
|
call (%rd255), _rt_buffer_get_64, (%rd256, %r109, %r381, %rd23, %rd24, %rd260, %rd260);
|
|
// inline asm
|
|
mov.f32 %f1220, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f1220;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs131, 0;
|
|
st.v4.u16 [%rd255], {%rs130, %rs130, %rs130, %rs131};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r382, [additive];
|
|
setp.eq.s32 %p154, %r382, 0;
|
|
@%p154 bra BB0_158;
|
|
|
|
mov.u64 %rd274, image_RNM1;
|
|
cvta.global.u64 %rd263, %rd274;
|
|
mov.u32 %r386, 8;
|
|
mov.u64 %rd273, 0;
|
|
// inline asm
|
|
call (%rd262), _rt_buffer_get_64, (%rd263, %r109, %r386, %rd23, %rd24, %rd273, %rd273);
|
|
// inline asm
|
|
ld.v4.u16 {%rs138, %rs139, %rs140, %rs141}, [%rd262];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1221, %rs138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1222, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1223, %rs140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd268), _rt_buffer_get_64, (%rd263, %r109, %r386, %rd23, %rd24, %rd273, %rd273);
|
|
// inline asm
|
|
add.f32 %f1224, %f1221, 0f00000000;
|
|
add.f32 %f1225, %f1222, 0f00000000;
|
|
add.f32 %f1226, %f1223, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1225;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f1224;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs142, 0;
|
|
st.v4.u16 [%rd268], {%rs135, %rs136, %rs137, %rs142};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd281, image_RNM1;
|
|
cvta.global.u64 %rd276, %rd281;
|
|
mov.u32 %r388, 8;
|
|
mov.u64 %rd280, 0;
|
|
// inline asm
|
|
call (%rd275), _rt_buffer_get_64, (%rd276, %r109, %r388, %rd23, %rd24, %rd280, %rd280);
|
|
// inline asm
|
|
mov.f32 %f1227, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1227;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs144, 0;
|
|
st.v4.u16 [%rd275], {%rs143, %rs143, %rs143, %rs144};
|
|
|
|
BB0_159:
|
|
ld.global.u32 %r389, [additive];
|
|
setp.eq.s32 %p155, %r389, 0;
|
|
@%p155 bra BB0_161;
|
|
|
|
mov.u64 %rd294, image_RNM2;
|
|
cvta.global.u64 %rd283, %rd294;
|
|
mov.u32 %r393, 8;
|
|
mov.u64 %rd293, 0;
|
|
// inline asm
|
|
call (%rd282), _rt_buffer_get_64, (%rd283, %r109, %r393, %rd23, %rd24, %rd293, %rd293);
|
|
// inline asm
|
|
ld.v4.u16 {%rs151, %rs152, %rs153, %rs154}, [%rd282];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1228, %rs151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1229, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1230, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd288), _rt_buffer_get_64, (%rd283, %r109, %r393, %rd23, %rd24, %rd293, %rd293);
|
|
// inline asm
|
|
add.f32 %f1231, %f1228, 0f00000000;
|
|
add.f32 %f1232, %f1229, 0f00000000;
|
|
add.f32 %f1233, %f1230, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1233;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1232;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f1231;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs155, 0;
|
|
st.v4.u16 [%rd288], {%rs148, %rs149, %rs150, %rs155};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_161:
|
|
mov.u64 %rd301, image_RNM2;
|
|
cvta.global.u64 %rd296, %rd301;
|
|
mov.u32 %r395, 8;
|
|
mov.u64 %rd300, 0;
|
|
// inline asm
|
|
call (%rd295), _rt_buffer_get_64, (%rd296, %r109, %r395, %rd23, %rd24, %rd300, %rd300);
|
|
// inline asm
|
|
mov.f32 %f1234, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f1234;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs157, 0;
|
|
st.v4.u16 [%rd295], {%rs156, %rs156, %rs156, %rs157};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_88:
|
|
setp.geu.f32 %p85, %f272, 0f00000000;
|
|
@%p85 bra BB0_91;
|
|
|
|
cvt.rzi.f32.f32 %f885, %f857;
|
|
setp.neu.f32 %p86, %f885, 0f3EE8BA2E;
|
|
selp.f32 %f1374, 0f7FFFFFFF, %f1374, %p86;
|
|
|
|
BB0_91:
|
|
add.f32 %f887, %f274, 0f3EE8BA2E;
|
|
mov.b32 %r296, %f887;
|
|
setp.lt.s32 %p88, %r296, 2139095040;
|
|
@%p88 bra BB0_96;
|
|
|
|
setp.gtu.f32 %p89, %f274, 0f7F800000;
|
|
@%p89 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f1374, %f272, 0f3EE8BA2E;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.neu.f32 %p90, %f274, 0f7F800000;
|
|
@%p90 bra BB0_96;
|
|
|
|
selp.f32 %f1374, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_96:
|
|
mul.f32 %f888, %f1374, 0f437F0000;
|
|
setp.eq.f32 %p91, %f272, 0f3F800000;
|
|
selp.f32 %f889, 0f437F0000, %f888, %p91;
|
|
cvt.rzi.u32.f32 %r297, %f889;
|
|
cvt.u16.u32 %rs41, %r297;
|
|
mov.u16 %rs42, 255;
|
|
st.v2.u8 [%rd114], {%rs41, %rs42};
|
|
ld.global.u32 %r429, [imageEnabled];
|
|
|
|
BB0_97:
|
|
ld.global.f32 %f890, [lightColor];
|
|
mul.f32 %f286, %f1368, %f890;
|
|
ld.global.f32 %f891, [lightColor+4];
|
|
mul.f32 %f287, %f1369, %f891;
|
|
ld.global.f32 %f892, [lightColor+8];
|
|
mul.f32 %f288, %f1370, %f892;
|
|
and.b32 %r298, %r429, 1;
|
|
setp.eq.b32 %p92, %r298, 1;
|
|
@!%p92 bra BB0_132;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_98:
|
|
mov.f32 %f895, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f896, %f895;
|
|
fma.rn.f32 %f897, %f896, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f289, %f897;
|
|
abs.f32 %f290, %f286;
|
|
setp.lt.f32 %p93, %f290, 0f00800000;
|
|
mul.f32 %f898, %f290, 0f4B800000;
|
|
selp.f32 %f899, 0fC3170000, 0fC2FE0000, %p93;
|
|
selp.f32 %f900, %f898, %f290, %p93;
|
|
mov.b32 %r299, %f900;
|
|
and.b32 %r300, %r299, 8388607;
|
|
or.b32 %r301, %r300, 1065353216;
|
|
mov.b32 %f901, %r301;
|
|
shr.u32 %r302, %r299, 23;
|
|
cvt.rn.f32.u32 %f902, %r302;
|
|
add.f32 %f903, %f899, %f902;
|
|
setp.gt.f32 %p94, %f901, 0f3FB504F3;
|
|
mul.f32 %f904, %f901, 0f3F000000;
|
|
add.f32 %f905, %f903, 0f3F800000;
|
|
selp.f32 %f906, %f904, %f901, %p94;
|
|
selp.f32 %f907, %f905, %f903, %p94;
|
|
add.f32 %f908, %f906, 0fBF800000;
|
|
add.f32 %f894, %f906, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f893,%f894;
|
|
// inline asm
|
|
add.f32 %f909, %f908, %f908;
|
|
mul.f32 %f910, %f893, %f909;
|
|
mul.f32 %f911, %f910, %f910;
|
|
mov.f32 %f912, 0f3C4CAF63;
|
|
mov.f32 %f913, 0f3B18F0FE;
|
|
fma.rn.f32 %f914, %f913, %f911, %f912;
|
|
mov.f32 %f915, 0f3DAAAABD;
|
|
fma.rn.f32 %f916, %f914, %f911, %f915;
|
|
mul.rn.f32 %f917, %f916, %f911;
|
|
mul.rn.f32 %f918, %f917, %f910;
|
|
sub.f32 %f919, %f908, %f910;
|
|
neg.f32 %f920, %f910;
|
|
add.f32 %f921, %f919, %f919;
|
|
fma.rn.f32 %f922, %f920, %f908, %f921;
|
|
mul.rn.f32 %f923, %f893, %f922;
|
|
add.f32 %f924, %f918, %f910;
|
|
sub.f32 %f925, %f910, %f924;
|
|
add.f32 %f926, %f918, %f925;
|
|
add.f32 %f927, %f923, %f926;
|
|
add.f32 %f928, %f924, %f927;
|
|
sub.f32 %f929, %f924, %f928;
|
|
add.f32 %f930, %f927, %f929;
|
|
mov.f32 %f931, 0f3F317200;
|
|
mul.rn.f32 %f932, %f907, %f931;
|
|
mov.f32 %f933, 0f35BFBE8E;
|
|
mul.rn.f32 %f934, %f907, %f933;
|
|
add.f32 %f935, %f932, %f928;
|
|
sub.f32 %f936, %f932, %f935;
|
|
add.f32 %f937, %f928, %f936;
|
|
add.f32 %f938, %f930, %f937;
|
|
add.f32 %f939, %f934, %f938;
|
|
add.f32 %f940, %f935, %f939;
|
|
sub.f32 %f941, %f935, %f940;
|
|
add.f32 %f942, %f939, %f941;
|
|
mov.f32 %f943, 0f3EE66666;
|
|
mul.rn.f32 %f944, %f943, %f940;
|
|
neg.f32 %f945, %f944;
|
|
fma.rn.f32 %f946, %f943, %f940, %f945;
|
|
fma.rn.f32 %f947, %f943, %f942, %f946;
|
|
mov.f32 %f948, 0f00000000;
|
|
fma.rn.f32 %f949, %f948, %f940, %f947;
|
|
add.rn.f32 %f950, %f944, %f949;
|
|
neg.f32 %f951, %f950;
|
|
add.rn.f32 %f952, %f944, %f951;
|
|
add.rn.f32 %f953, %f952, %f949;
|
|
mov.b32 %r303, %f950;
|
|
setp.eq.s32 %p95, %r303, 1118925336;
|
|
add.s32 %r304, %r303, -1;
|
|
mov.b32 %f954, %r304;
|
|
add.f32 %f955, %f953, 0f37000000;
|
|
selp.f32 %f956, %f954, %f950, %p95;
|
|
selp.f32 %f291, %f955, %f953, %p95;
|
|
mul.f32 %f957, %f956, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f958, %f957;
|
|
mov.f32 %f959, 0fBF317200;
|
|
fma.rn.f32 %f960, %f958, %f959, %f956;
|
|
mov.f32 %f961, 0fB5BFBE8E;
|
|
fma.rn.f32 %f962, %f958, %f961, %f960;
|
|
mul.f32 %f963, %f962, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f964, %f963;
|
|
add.f32 %f965, %f958, 0f00000000;
|
|
ex2.approx.f32 %f966, %f965;
|
|
mul.f32 %f967, %f964, %f966;
|
|
setp.lt.f32 %p96, %f956, 0fC2D20000;
|
|
selp.f32 %f968, 0f00000000, %f967, %p96;
|
|
setp.gt.f32 %p97, %f956, 0f42D20000;
|
|
selp.f32 %f1375, 0f7F800000, %f968, %p97;
|
|
setp.eq.f32 %p98, %f1375, 0f7F800000;
|
|
@%p98 bra BB0_100;
|
|
|
|
fma.rn.f32 %f1375, %f1375, %f291, %f1375;
|
|
|
|
BB0_100:
|
|
setp.lt.f32 %p99, %f286, 0f00000000;
|
|
setp.eq.f32 %p100, %f289, 0f3F800000;
|
|
and.pred %p4, %p99, %p100;
|
|
mov.b32 %r305, %f1375;
|
|
xor.b32 %r306, %r305, -2147483648;
|
|
mov.b32 %f969, %r306;
|
|
selp.f32 %f1377, %f969, %f1375, %p4;
|
|
setp.eq.f32 %p101, %f286, 0f00000000;
|
|
@%p101 bra BB0_103;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_103:
|
|
add.f32 %f972, %f286, %f286;
|
|
selp.f32 %f1377, %f972, 0f00000000, %p100;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_101:
|
|
setp.geu.f32 %p102, %f286, 0f00000000;
|
|
@%p102 bra BB0_104;
|
|
|
|
cvt.rzi.f32.f32 %f971, %f943;
|
|
setp.neu.f32 %p103, %f971, 0f3EE66666;
|
|
selp.f32 %f1377, 0f7FFFFFFF, %f1377, %p103;
|
|
|
|
BB0_104:
|
|
add.f32 %f973, %f290, 0f3EE66666;
|
|
mov.b32 %r307, %f973;
|
|
setp.lt.s32 %p105, %r307, 2139095040;
|
|
@%p105 bra BB0_109;
|
|
|
|
setp.gtu.f32 %p106, %f290, 0f7F800000;
|
|
@%p106 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f1377, %f286, 0f3EE66666;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.neu.f32 %p107, %f290, 0f7F800000;
|
|
@%p107 bra BB0_109;
|
|
|
|
selp.f32 %f1377, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_109:
|
|
setp.eq.f32 %p108, %f286, 0f3F800000;
|
|
selp.f32 %f302, 0f3F800000, %f1377, %p108;
|
|
abs.f32 %f303, %f287;
|
|
setp.lt.f32 %p109, %f303, 0f00800000;
|
|
mul.f32 %f976, %f303, 0f4B800000;
|
|
selp.f32 %f977, 0fC3170000, 0fC2FE0000, %p109;
|
|
selp.f32 %f978, %f976, %f303, %p109;
|
|
mov.b32 %r308, %f978;
|
|
and.b32 %r309, %r308, 8388607;
|
|
or.b32 %r310, %r309, 1065353216;
|
|
mov.b32 %f979, %r310;
|
|
shr.u32 %r311, %r308, 23;
|
|
cvt.rn.f32.u32 %f980, %r311;
|
|
add.f32 %f981, %f977, %f980;
|
|
setp.gt.f32 %p110, %f979, 0f3FB504F3;
|
|
mul.f32 %f982, %f979, 0f3F000000;
|
|
add.f32 %f983, %f981, 0f3F800000;
|
|
selp.f32 %f984, %f982, %f979, %p110;
|
|
selp.f32 %f985, %f983, %f981, %p110;
|
|
add.f32 %f986, %f984, 0fBF800000;
|
|
add.f32 %f975, %f984, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f974,%f975;
|
|
// inline asm
|
|
add.f32 %f987, %f986, %f986;
|
|
mul.f32 %f988, %f974, %f987;
|
|
mul.f32 %f989, %f988, %f988;
|
|
fma.rn.f32 %f992, %f913, %f989, %f912;
|
|
fma.rn.f32 %f994, %f992, %f989, %f915;
|
|
mul.rn.f32 %f995, %f994, %f989;
|
|
mul.rn.f32 %f996, %f995, %f988;
|
|
sub.f32 %f997, %f986, %f988;
|
|
neg.f32 %f998, %f988;
|
|
add.f32 %f999, %f997, %f997;
|
|
fma.rn.f32 %f1000, %f998, %f986, %f999;
|
|
mul.rn.f32 %f1001, %f974, %f1000;
|
|
add.f32 %f1002, %f996, %f988;
|
|
sub.f32 %f1003, %f988, %f1002;
|
|
add.f32 %f1004, %f996, %f1003;
|
|
add.f32 %f1005, %f1001, %f1004;
|
|
add.f32 %f1006, %f1002, %f1005;
|
|
sub.f32 %f1007, %f1002, %f1006;
|
|
add.f32 %f1008, %f1005, %f1007;
|
|
mul.rn.f32 %f1010, %f985, %f931;
|
|
mul.rn.f32 %f1012, %f985, %f933;
|
|
add.f32 %f1013, %f1010, %f1006;
|
|
sub.f32 %f1014, %f1010, %f1013;
|
|
add.f32 %f1015, %f1006, %f1014;
|
|
add.f32 %f1016, %f1008, %f1015;
|
|
add.f32 %f1017, %f1012, %f1016;
|
|
add.f32 %f1018, %f1013, %f1017;
|
|
sub.f32 %f1019, %f1013, %f1018;
|
|
add.f32 %f1020, %f1017, %f1019;
|
|
mul.rn.f32 %f1022, %f943, %f1018;
|
|
neg.f32 %f1023, %f1022;
|
|
fma.rn.f32 %f1024, %f943, %f1018, %f1023;
|
|
fma.rn.f32 %f1025, %f943, %f1020, %f1024;
|
|
fma.rn.f32 %f1027, %f948, %f1018, %f1025;
|
|
add.rn.f32 %f1028, %f1022, %f1027;
|
|
neg.f32 %f1029, %f1028;
|
|
add.rn.f32 %f1030, %f1022, %f1029;
|
|
add.rn.f32 %f1031, %f1030, %f1027;
|
|
mov.b32 %r312, %f1028;
|
|
setp.eq.s32 %p111, %r312, 1118925336;
|
|
add.s32 %r313, %r312, -1;
|
|
mov.b32 %f1032, %r313;
|
|
add.f32 %f1033, %f1031, 0f37000000;
|
|
selp.f32 %f1034, %f1032, %f1028, %p111;
|
|
selp.f32 %f304, %f1033, %f1031, %p111;
|
|
mul.f32 %f1035, %f1034, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1036, %f1035;
|
|
fma.rn.f32 %f1038, %f1036, %f959, %f1034;
|
|
fma.rn.f32 %f1040, %f1036, %f961, %f1038;
|
|
mul.f32 %f1041, %f1040, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1042, %f1041;
|
|
add.f32 %f1043, %f1036, 0f00000000;
|
|
ex2.approx.f32 %f1044, %f1043;
|
|
mul.f32 %f1045, %f1042, %f1044;
|
|
setp.lt.f32 %p112, %f1034, 0fC2D20000;
|
|
selp.f32 %f1046, 0f00000000, %f1045, %p112;
|
|
setp.gt.f32 %p113, %f1034, 0f42D20000;
|
|
selp.f32 %f1378, 0f7F800000, %f1046, %p113;
|
|
setp.eq.f32 %p114, %f1378, 0f7F800000;
|
|
@%p114 bra BB0_111;
|
|
|
|
fma.rn.f32 %f1378, %f1378, %f304, %f1378;
|
|
|
|
BB0_111:
|
|
setp.lt.f32 %p115, %f287, 0f00000000;
|
|
and.pred %p5, %p115, %p100;
|
|
mov.b32 %r314, %f1378;
|
|
xor.b32 %r315, %r314, -2147483648;
|
|
mov.b32 %f1047, %r315;
|
|
selp.f32 %f1380, %f1047, %f1378, %p5;
|
|
setp.eq.f32 %p117, %f287, 0f00000000;
|
|
@%p117 bra BB0_114;
|
|
bra.uni BB0_112;
|
|
|
|
BB0_114:
|
|
add.f32 %f1050, %f287, %f287;
|
|
selp.f32 %f1380, %f1050, 0f00000000, %p100;
|
|
bra.uni BB0_115;
|
|
|
|
BB0_112:
|
|
setp.geu.f32 %p118, %f287, 0f00000000;
|
|
@%p118 bra BB0_115;
|
|
|
|
cvt.rzi.f32.f32 %f1049, %f943;
|
|
setp.neu.f32 %p119, %f1049, 0f3EE66666;
|
|
selp.f32 %f1380, 0f7FFFFFFF, %f1380, %p119;
|
|
|
|
BB0_115:
|
|
add.f32 %f1051, %f303, 0f3EE66666;
|
|
mov.b32 %r316, %f1051;
|
|
setp.lt.s32 %p121, %r316, 2139095040;
|
|
@%p121 bra BB0_120;
|
|
|
|
setp.gtu.f32 %p122, %f303, 0f7F800000;
|
|
@%p122 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f1380, %f287, 0f3EE66666;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.neu.f32 %p123, %f303, 0f7F800000;
|
|
@%p123 bra BB0_120;
|
|
|
|
selp.f32 %f1380, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_120:
|
|
setp.eq.f32 %p124, %f287, 0f3F800000;
|
|
selp.f32 %f315, 0f3F800000, %f1380, %p124;
|
|
abs.f32 %f316, %f288;
|
|
setp.lt.f32 %p125, %f316, 0f00800000;
|
|
mul.f32 %f1054, %f316, 0f4B800000;
|
|
selp.f32 %f1055, 0fC3170000, 0fC2FE0000, %p125;
|
|
selp.f32 %f1056, %f1054, %f316, %p125;
|
|
mov.b32 %r317, %f1056;
|
|
and.b32 %r318, %r317, 8388607;
|
|
or.b32 %r319, %r318, 1065353216;
|
|
mov.b32 %f1057, %r319;
|
|
shr.u32 %r320, %r317, 23;
|
|
cvt.rn.f32.u32 %f1058, %r320;
|
|
add.f32 %f1059, %f1055, %f1058;
|
|
setp.gt.f32 %p126, %f1057, 0f3FB504F3;
|
|
mul.f32 %f1060, %f1057, 0f3F000000;
|
|
add.f32 %f1061, %f1059, 0f3F800000;
|
|
selp.f32 %f1062, %f1060, %f1057, %p126;
|
|
selp.f32 %f1063, %f1061, %f1059, %p126;
|
|
add.f32 %f1064, %f1062, 0fBF800000;
|
|
add.f32 %f1053, %f1062, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1052,%f1053;
|
|
// inline asm
|
|
add.f32 %f1065, %f1064, %f1064;
|
|
mul.f32 %f1066, %f1052, %f1065;
|
|
mul.f32 %f1067, %f1066, %f1066;
|
|
fma.rn.f32 %f1070, %f913, %f1067, %f912;
|
|
fma.rn.f32 %f1072, %f1070, %f1067, %f915;
|
|
mul.rn.f32 %f1073, %f1072, %f1067;
|
|
mul.rn.f32 %f1074, %f1073, %f1066;
|
|
sub.f32 %f1075, %f1064, %f1066;
|
|
neg.f32 %f1076, %f1066;
|
|
add.f32 %f1077, %f1075, %f1075;
|
|
fma.rn.f32 %f1078, %f1076, %f1064, %f1077;
|
|
mul.rn.f32 %f1079, %f1052, %f1078;
|
|
add.f32 %f1080, %f1074, %f1066;
|
|
sub.f32 %f1081, %f1066, %f1080;
|
|
add.f32 %f1082, %f1074, %f1081;
|
|
add.f32 %f1083, %f1079, %f1082;
|
|
add.f32 %f1084, %f1080, %f1083;
|
|
sub.f32 %f1085, %f1080, %f1084;
|
|
add.f32 %f1086, %f1083, %f1085;
|
|
mul.rn.f32 %f1088, %f1063, %f931;
|
|
mul.rn.f32 %f1090, %f1063, %f933;
|
|
add.f32 %f1091, %f1088, %f1084;
|
|
sub.f32 %f1092, %f1088, %f1091;
|
|
add.f32 %f1093, %f1084, %f1092;
|
|
add.f32 %f1094, %f1086, %f1093;
|
|
add.f32 %f1095, %f1090, %f1094;
|
|
add.f32 %f1096, %f1091, %f1095;
|
|
sub.f32 %f1097, %f1091, %f1096;
|
|
add.f32 %f1098, %f1095, %f1097;
|
|
mul.rn.f32 %f1100, %f943, %f1096;
|
|
neg.f32 %f1101, %f1100;
|
|
fma.rn.f32 %f1102, %f943, %f1096, %f1101;
|
|
fma.rn.f32 %f1103, %f943, %f1098, %f1102;
|
|
fma.rn.f32 %f1105, %f948, %f1096, %f1103;
|
|
add.rn.f32 %f1106, %f1100, %f1105;
|
|
neg.f32 %f1107, %f1106;
|
|
add.rn.f32 %f1108, %f1100, %f1107;
|
|
add.rn.f32 %f1109, %f1108, %f1105;
|
|
mov.b32 %r321, %f1106;
|
|
setp.eq.s32 %p127, %r321, 1118925336;
|
|
add.s32 %r322, %r321, -1;
|
|
mov.b32 %f1110, %r322;
|
|
add.f32 %f1111, %f1109, 0f37000000;
|
|
selp.f32 %f1112, %f1110, %f1106, %p127;
|
|
selp.f32 %f317, %f1111, %f1109, %p127;
|
|
mul.f32 %f1113, %f1112, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1114, %f1113;
|
|
fma.rn.f32 %f1116, %f1114, %f959, %f1112;
|
|
fma.rn.f32 %f1118, %f1114, %f961, %f1116;
|
|
mul.f32 %f1119, %f1118, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1120, %f1119;
|
|
add.f32 %f1121, %f1114, 0f00000000;
|
|
ex2.approx.f32 %f1122, %f1121;
|
|
mul.f32 %f1123, %f1120, %f1122;
|
|
setp.lt.f32 %p128, %f1112, 0fC2D20000;
|
|
selp.f32 %f1124, 0f00000000, %f1123, %p128;
|
|
setp.gt.f32 %p129, %f1112, 0f42D20000;
|
|
selp.f32 %f1381, 0f7F800000, %f1124, %p129;
|
|
setp.eq.f32 %p130, %f1381, 0f7F800000;
|
|
@%p130 bra BB0_122;
|
|
|
|
fma.rn.f32 %f1381, %f1381, %f317, %f1381;
|
|
|
|
BB0_122:
|
|
setp.lt.f32 %p131, %f288, 0f00000000;
|
|
and.pred %p6, %p131, %p100;
|
|
mov.b32 %r323, %f1381;
|
|
xor.b32 %r324, %r323, -2147483648;
|
|
mov.b32 %f1125, %r324;
|
|
selp.f32 %f1383, %f1125, %f1381, %p6;
|
|
setp.eq.f32 %p133, %f288, 0f00000000;
|
|
@%p133 bra BB0_125;
|
|
bra.uni BB0_123;
|
|
|
|
BB0_125:
|
|
add.f32 %f1128, %f288, %f288;
|
|
selp.f32 %f1383, %f1128, 0f00000000, %p100;
|
|
bra.uni BB0_126;
|
|
|
|
BB0_123:
|
|
setp.geu.f32 %p134, %f288, 0f00000000;
|
|
@%p134 bra BB0_126;
|
|
|
|
cvt.rzi.f32.f32 %f1127, %f943;
|
|
setp.neu.f32 %p135, %f1127, 0f3EE66666;
|
|
selp.f32 %f1383, 0f7FFFFFFF, %f1383, %p135;
|
|
|
|
BB0_126:
|
|
add.f32 %f1129, %f316, 0f3EE66666;
|
|
mov.b32 %r325, %f1129;
|
|
setp.lt.s32 %p137, %r325, 2139095040;
|
|
@%p137 bra BB0_131;
|
|
|
|
setp.gtu.f32 %p138, %f316, 0f7F800000;
|
|
@%p138 bra BB0_130;
|
|
bra.uni BB0_128;
|
|
|
|
BB0_130:
|
|
add.f32 %f1383, %f288, 0f3EE66666;
|
|
bra.uni BB0_131;
|
|
|
|
BB0_128:
|
|
setp.neu.f32 %p139, %f316, 0f7F800000;
|
|
@%p139 bra BB0_131;
|
|
|
|
selp.f32 %f1383, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_131:
|
|
setp.eq.f32 %p140, %f288, 0f3F800000;
|
|
selp.f32 %f1130, 0f3F800000, %f1383, %p140;
|
|
cvt.u64.u32 %rd124, %r6;
|
|
cvt.u64.u32 %rd123, %r5;
|
|
mov.u64 %rd127, image;
|
|
cvta.global.u64 %rd122, %rd127;
|
|
// inline asm
|
|
call (%rd121), _rt_buffer_get_64, (%rd122, %r109, %r110, %rd123, %rd124, %rd30, %rd30);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1131, %f1130;
|
|
mul.f32 %f1132, %f1131, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r328, %f1132;
|
|
cvt.sat.f32.f32 %f1133, %f315;
|
|
mul.f32 %f1134, %f1133, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r329, %f1134;
|
|
cvt.sat.f32.f32 %f1135, %f302;
|
|
mul.f32 %f1136, %f1135, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r330, %f1136;
|
|
cvt.u16.u32 %rs43, %r328;
|
|
cvt.u16.u32 %rs44, %r330;
|
|
cvt.u16.u32 %rs45, %r329;
|
|
mov.u16 %rs46, 255;
|
|
st.v4.u8 [%rd121], {%rs43, %rs45, %rs44, %rs46};
|
|
ld.global.u32 %r429, [imageEnabled];
|
|
|
|
BB0_132:
|
|
cvt.u64.u32 %rd21, %r5;
|
|
cvt.u64.u32 %rd22, %r6;
|
|
and.b32 %r331, %r429, 4;
|
|
setp.eq.s32 %p141, %r331, 0;
|
|
@%p141 bra BB0_136;
|
|
|
|
ld.global.u32 %r332, [additive];
|
|
setp.eq.s32 %p142, %r332, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f655;}
|
|
|
|
// inline asm
|
|
@%p142 bra BB0_135;
|
|
|
|
mov.u64 %rd140, image_HDR;
|
|
cvta.global.u64 %rd129, %rd140;
|
|
mov.u32 %r336, 8;
|
|
// inline asm
|
|
call (%rd128), _rt_buffer_get_64, (%rd129, %r109, %r336, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
ld.v4.u16 {%rs54, %rs55, %rs56, %rs57}, [%rd128];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1138, %rs54;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1139, %rs55;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1140, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd134), _rt_buffer_get_64, (%rd129, %r109, %r336, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
add.f32 %f1141, %f286, %f1138;
|
|
add.f32 %f1142, %f287, %f1139;
|
|
add.f32 %f1143, %f288, %f1140;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f1142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1141;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd134], {%rs51, %rs52, %rs53, %rs47};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd147, image_HDR;
|
|
cvta.global.u64 %rd142, %rd147;
|
|
mov.u32 %r338, 8;
|
|
// inline asm
|
|
call (%rd141), _rt_buffer_get_64, (%rd142, %r109, %r338, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f288;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f287;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f286;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd141], {%rs58, %rs59, %rs60, %rs47};
|
|
|
|
BB0_136:
|
|
ld.global.f32 %f1148, [lightColor];
|
|
mul.f32 %f1149, %f1365, %f1148;
|
|
ld.global.f32 %f1150, [lightColor+4];
|
|
mul.f32 %f1151, %f1366, %f1150;
|
|
ld.global.f32 %f1152, [lightColor+8];
|
|
mul.f32 %f1153, %f1367, %f1152;
|
|
mul.f32 %f1154, %f1362, %f1148;
|
|
mul.f32 %f1155, %f1363, %f1150;
|
|
mul.f32 %f1156, %f1364, %f1152;
|
|
mul.f32 %f1157, %f1359, %f1148;
|
|
mul.f32 %f1158, %f1360, %f1150;
|
|
mul.f32 %f1159, %f1361, %f1152;
|
|
add.f32 %f1160, %f1149, %f1154;
|
|
add.f32 %f1161, %f1151, %f1155;
|
|
add.f32 %f1162, %f1153, %f1156;
|
|
add.f32 %f1163, %f1160, %f1157;
|
|
add.f32 %f1164, %f1161, %f1158;
|
|
add.f32 %f1165, %f1162, %f1159;
|
|
mul.f32 %f1166, %f1163, 0f3F13CD3A;
|
|
mul.f32 %f1167, %f1164, 0f3F13CD3A;
|
|
mul.f32 %f1168, %f1165, 0f3F13CD3A;
|
|
div.rn.f32 %f1169, %f286, %f1166;
|
|
div.rn.f32 %f1170, %f287, %f1167;
|
|
div.rn.f32 %f1171, %f288, %f1168;
|
|
setp.eq.f32 %p143, %f286, 0f00000000;
|
|
selp.f32 %f1172, 0f00000000, %f1169, %p143;
|
|
setp.eq.f32 %p144, %f287, 0f00000000;
|
|
selp.f32 %f1173, 0f00000000, %f1170, %p144;
|
|
setp.eq.f32 %p145, %f288, 0f00000000;
|
|
selp.f32 %f1174, 0f00000000, %f1171, %p145;
|
|
mul.f32 %f328, %f1149, %f1172;
|
|
mul.f32 %f329, %f1151, %f1173;
|
|
mul.f32 %f330, %f1153, %f1174;
|
|
mul.f32 %f331, %f1154, %f1172;
|
|
mul.f32 %f332, %f1155, %f1173;
|
|
mul.f32 %f333, %f1156, %f1174;
|
|
mul.f32 %f334, %f1157, %f1172;
|
|
mul.f32 %f335, %f1158, %f1173;
|
|
mul.f32 %f336, %f1159, %f1174;
|
|
ld.global.u32 %r339, [additive];
|
|
setp.eq.s32 %p146, %r339, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f655;}
|
|
|
|
// inline asm
|
|
@%p146 bra BB0_138;
|
|
|
|
mov.u64 %rd160, image_RNM0;
|
|
cvta.global.u64 %rd149, %rd160;
|
|
mov.u32 %r343, 8;
|
|
// inline asm
|
|
call (%rd148), _rt_buffer_get_64, (%rd149, %r109, %r343, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
ld.v4.u16 {%rs68, %rs69, %rs70, %rs71}, [%rd148];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1175, %rs68;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1176, %rs69;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1177, %rs70;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd154), _rt_buffer_get_64, (%rd149, %r109, %r343, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
add.f32 %f1178, %f328, %f1175;
|
|
add.f32 %f1179, %f329, %f1176;
|
|
add.f32 %f1180, %f330, %f1177;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f1180;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f1179;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f1178;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd154], {%rs65, %rs66, %rs67, %rs61};
|
|
bra.uni BB0_139;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd167, image_RNM0;
|
|
cvta.global.u64 %rd162, %rd167;
|
|
mov.u32 %r345, 8;
|
|
// inline asm
|
|
call (%rd161), _rt_buffer_get_64, (%rd162, %r109, %r345, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f330;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f329;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f328;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd161], {%rs72, %rs73, %rs74, %rs61};
|
|
|
|
BB0_139:
|
|
ld.global.u32 %r346, [additive];
|
|
setp.eq.s32 %p147, %r346, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f655;}
|
|
|
|
// inline asm
|
|
@%p147 bra BB0_141;
|
|
|
|
mov.u64 %rd180, image_RNM1;
|
|
cvta.global.u64 %rd169, %rd180;
|
|
mov.u32 %r350, 8;
|
|
// inline asm
|
|
call (%rd168), _rt_buffer_get_64, (%rd169, %r109, %r350, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
ld.v4.u16 {%rs82, %rs83, %rs84, %rs85}, [%rd168];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1185, %rs82;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1186, %rs83;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1187, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd174), _rt_buffer_get_64, (%rd169, %r109, %r350, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
add.f32 %f1188, %f331, %f1185;
|
|
add.f32 %f1189, %f332, %f1186;
|
|
add.f32 %f1190, %f333, %f1187;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1190;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f1189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1188;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd174], {%rs79, %rs80, %rs81, %rs75};
|
|
bra.uni BB0_142;
|
|
|
|
BB0_141:
|
|
mov.u64 %rd187, image_RNM1;
|
|
cvta.global.u64 %rd182, %rd187;
|
|
mov.u32 %r352, 8;
|
|
// inline asm
|
|
call (%rd181), _rt_buffer_get_64, (%rd182, %r109, %r352, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f333;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f332;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f331;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd181], {%rs86, %rs87, %rs88, %rs75};
|
|
|
|
BB0_142:
|
|
ld.global.u32 %r353, [additive];
|
|
setp.eq.s32 %p148, %r353, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f655;}
|
|
|
|
// inline asm
|
|
@%p148 bra BB0_144;
|
|
|
|
mov.u64 %rd200, image_RNM2;
|
|
cvta.global.u64 %rd189, %rd200;
|
|
mov.u32 %r357, 8;
|
|
// inline asm
|
|
call (%rd188), _rt_buffer_get_64, (%rd189, %r109, %r357, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
ld.v4.u16 {%rs96, %rs97, %rs98, %rs99}, [%rd188];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1195, %rs96;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1196, %rs97;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1197, %rs98;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd194), _rt_buffer_get_64, (%rd189, %r109, %r357, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
add.f32 %f1198, %f334, %f1195;
|
|
add.f32 %f1199, %f335, %f1196;
|
|
add.f32 %f1200, %f336, %f1197;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f1200;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f1199;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f1198;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd194], {%rs93, %rs94, %rs95, %rs89};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_144:
|
|
mov.u64 %rd207, image_RNM2;
|
|
cvta.global.u64 %rd202, %rd207;
|
|
mov.u32 %r359, 8;
|
|
// inline asm
|
|
call (%rd201), _rt_buffer_get_64, (%rd202, %r109, %r359, %rd21, %rd22, %rd30, %rd30);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f336;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f335;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f334;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd201], {%rs100, %rs101, %rs102, %rs89};
|
|
|
|
BB0_162:
|
|
ret;
|
|
}
|
|
|
|
|