2484 lines
70 KiB
Plaintext
2484 lines
70 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 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[36];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<156>;
|
|
.reg .b16 %rs<157>;
|
|
.reg .f32 %f<1168>;
|
|
.reg .b32 %r<435>;
|
|
.reg .b64 %rd<292>;
|
|
|
|
|
|
mov.u64 %rd291, __local_depot0;
|
|
cvta.local.u64 %SP, %rd291;
|
|
ld.global.u32 %r1, [samples];
|
|
ld.global.v2.u32 {%r109, %r110}, [pixelID];
|
|
cvt.u64.u32 %rd26, %r109;
|
|
cvt.u64.u32 %rd27, %r110;
|
|
mov.u64 %rd30, uvnormal;
|
|
cvta.global.u64 %rd25, %rd30;
|
|
mov.u32 %r107, 2;
|
|
mov.u32 %r108, 4;
|
|
mov.u64 %rd29, 0;
|
|
// inline asm
|
|
call (%rd24), _rt_buffer_get_64, (%rd25, %r107, %r108, %rd26, %rd27, %rd29, %rd29);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd24];
|
|
shr.u32 %r113, %r2, 16;
|
|
cvt.u16.u32 %rs1, %r113;
|
|
and.b16 %rs6, %rs1, 255;
|
|
cvt.u16.u32 %rs7, %r2;
|
|
or.b16 %rs8, %rs7, %rs6;
|
|
setp.eq.s16 %p7, %rs8, 0;
|
|
mov.f32 %f1089, 0f00000000;
|
|
mov.f32 %f1090, %f1089;
|
|
mov.f32 %f1091, %f1089;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs9, [%rd24+1];
|
|
and.b16 %rs11, %rs7, 255;
|
|
cvt.rn.f32.u16 %f248, %rs11;
|
|
div.rn.f32 %f249, %f248, 0f437F0000;
|
|
fma.rn.f32 %f250, %f249, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f251, %rs9;
|
|
div.rn.f32 %f252, %f251, 0f437F0000;
|
|
fma.rn.f32 %f253, %f252, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f254, %rs6;
|
|
div.rn.f32 %f255, %f254, 0f437F0000;
|
|
fma.rn.f32 %f256, %f255, 0f40000000, 0fBF800000;
|
|
mul.f32 %f257, %f253, %f253;
|
|
fma.rn.f32 %f258, %f250, %f250, %f257;
|
|
fma.rn.f32 %f259, %f256, %f256, %f258;
|
|
sqrt.rn.f32 %f260, %f259;
|
|
rcp.rn.f32 %f261, %f260;
|
|
mul.f32 %f1089, %f250, %f261;
|
|
mul.f32 %f1090, %f253, %f261;
|
|
mul.f32 %f1091, %f256, %f261;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r114, %r115}, [pixelID];
|
|
ld.global.v2.u32 {%r117, %r118}, [tileInfo];
|
|
add.s32 %r3, %r114, %r117;
|
|
add.s32 %r4, %r115, %r118;
|
|
setp.eq.f32 %p8, %f1090, 0f00000000;
|
|
setp.eq.f32 %p9, %f1089, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1091, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_144;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_144:
|
|
ld.global.u32 %r434, [imageEnabled];
|
|
and.b32 %r353, %r434, 1;
|
|
setp.eq.b32 %p148, %r353, 1;
|
|
@!%p148 bra BB0_146;
|
|
bra.uni BB0_145;
|
|
|
|
BB0_145:
|
|
cvt.u64.u32 %rd193, %r3;
|
|
cvt.u64.u32 %rd194, %r4;
|
|
mov.u64 %rd197, image;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u64 %rd196, 0;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r107, %r108, %rd193, %rd194, %rd196, %rd196);
|
|
// inline asm
|
|
mov.u16 %rs102, 0;
|
|
st.v4.u8 [%rd191], {%rs102, %rs102, %rs102, %rs102};
|
|
ld.global.u32 %r434, [imageEnabled];
|
|
|
|
BB0_146:
|
|
and.b32 %r356, %r434, 8;
|
|
setp.eq.s32 %p149, %r356, 0;
|
|
@%p149 bra BB0_148;
|
|
|
|
cvt.u64.u32 %rd201, %r4;
|
|
cvt.u64.u32 %rd200, %r3;
|
|
mov.u64 %rd204, image_Mask;
|
|
cvta.global.u64 %rd199, %rd204;
|
|
mov.u64 %rd203, 0;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r107, %r107, %rd200, %rd201, %rd203, %rd203);
|
|
// inline asm
|
|
mov.f32 %f1051, 0f00000000;
|
|
cvt.rzi.u32.f32 %r359, %f1051;
|
|
cvt.u16.u32 %rs103, %r359;
|
|
mov.u16 %rs104, 0;
|
|
st.v2.u8 [%rd198], {%rs103, %rs104};
|
|
ld.global.u32 %r434, [imageEnabled];
|
|
|
|
BB0_148:
|
|
cvt.u64.u32 %rd22, %r3;
|
|
cvt.u64.u32 %rd23, %r4;
|
|
and.b32 %r360, %r434, 4;
|
|
setp.eq.s32 %p150, %r360, 0;
|
|
@%p150 bra BB0_152;
|
|
|
|
ld.global.u32 %r361, [additive];
|
|
setp.eq.s32 %p151, %r361, 0;
|
|
@%p151 bra BB0_151;
|
|
|
|
mov.u64 %rd217, image_HDR;
|
|
cvta.global.u64 %rd206, %rd217;
|
|
mov.u32 %r365, 8;
|
|
mov.u64 %rd216, 0;
|
|
// inline asm
|
|
call (%rd205), _rt_buffer_get_64, (%rd206, %r107, %r365, %rd22, %rd23, %rd216, %rd216);
|
|
// inline asm
|
|
ld.v4.u16 {%rs111, %rs112, %rs113, %rs114}, [%rd205];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1052, %rs111;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1053, %rs112;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1054, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd206, %r107, %r365, %rd22, %rd23, %rd216, %rd216);
|
|
// inline asm
|
|
add.f32 %f1055, %f1052, 0f00000000;
|
|
add.f32 %f1056, %f1053, 0f00000000;
|
|
add.f32 %f1057, %f1054, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1057;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f1056;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f1055;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs115, 0;
|
|
st.v4.u16 [%rd211], {%rs108, %rs109, %rs110, %rs115};
|
|
bra.uni BB0_152;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r126, %r127}, [pixelID];
|
|
cvt.u64.u32 %rd33, %r126;
|
|
cvt.u64.u32 %rd34, %r127;
|
|
mov.u64 %rd43, uvpos;
|
|
cvta.global.u64 %rd32, %rd43;
|
|
mov.u32 %r123, 12;
|
|
// inline asm
|
|
call (%rd31), _rt_buffer_get_64, (%rd32, %r107, %r123, %rd33, %rd34, %rd29, %rd29);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd31+8];
|
|
ld.f32 %f8, [%rd31+4];
|
|
ld.f32 %f7, [%rd31];
|
|
mul.f32 %f265, %f7, 0f3456BF95;
|
|
mul.f32 %f266, %f8, 0f3456BF95;
|
|
mul.f32 %f267, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f1089;
|
|
div.rn.f32 %f268, %f265, %f10;
|
|
abs.f32 %f269, %f1090;
|
|
div.rn.f32 %f270, %f266, %f269;
|
|
abs.f32 %f11, %f1091;
|
|
div.rn.f32 %f271, %f267, %f11;
|
|
abs.f32 %f272, %f268;
|
|
abs.f32 %f273, %f270;
|
|
abs.f32 %f274, %f271;
|
|
mov.f32 %f275, 0f38D1B717;
|
|
max.f32 %f276, %f272, %f275;
|
|
max.f32 %f277, %f273, %f275;
|
|
max.f32 %f278, %f274, %f275;
|
|
fma.rn.f32 %f12, %f1089, %f276, %f7;
|
|
fma.rn.f32 %f13, %f1090, %f277, %f8;
|
|
fma.rn.f32 %f14, %f1091, %f278, %f9;
|
|
ld.global.v2.u32 {%r130, %r131}, [pixelID];
|
|
cvt.u64.u32 %rd39, %r130;
|
|
cvt.u64.u32 %rd40, %r131;
|
|
mov.u64 %rd44, uvtangent;
|
|
cvta.global.u64 %rd38, %rd44;
|
|
// inline asm
|
|
call (%rd37), _rt_buffer_get_64, (%rd38, %r107, %r108, %rd39, %rd40, %rd29, %rd29);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd37];
|
|
shr.u32 %r6, %r5, 16;
|
|
cvt.u16.u32 %rs13, %r6;
|
|
and.b16 %rs14, %rs13, 255;
|
|
cvt.u16.u32 %rs15, %r5;
|
|
or.b16 %rs16, %rs15, %rs14;
|
|
setp.eq.s16 %p13, %rs16, 0;
|
|
mov.f32 %f50, 0f00000000;
|
|
mov.f32 %f1092, %f50;
|
|
mov.f32 %f1093, %f50;
|
|
mov.f32 %f1094, %f50;
|
|
@%p13 bra BB0_5;
|
|
|
|
ld.u8 %rs17, [%rd37+1];
|
|
and.b16 %rs19, %rs15, 255;
|
|
cvt.rn.f32.u16 %f279, %rs19;
|
|
div.rn.f32 %f280, %f279, 0f437F0000;
|
|
fma.rn.f32 %f281, %f280, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f282, %rs17;
|
|
div.rn.f32 %f283, %f282, 0f437F0000;
|
|
fma.rn.f32 %f284, %f283, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f285, %rs14;
|
|
div.rn.f32 %f286, %f285, 0f437F0000;
|
|
fma.rn.f32 %f287, %f286, 0f40000000, 0fBF800000;
|
|
mul.f32 %f288, %f284, %f284;
|
|
fma.rn.f32 %f289, %f281, %f281, %f288;
|
|
fma.rn.f32 %f290, %f287, %f287, %f289;
|
|
sqrt.rn.f32 %f291, %f290;
|
|
rcp.rn.f32 %f292, %f291;
|
|
mul.f32 %f1092, %f281, %f292;
|
|
mul.f32 %f1093, %f284, %f292;
|
|
mul.f32 %f1094, %f287, %f292;
|
|
|
|
BB0_5:
|
|
mul.f32 %f296, %f1091, %f1093;
|
|
mul.f32 %f297, %f1090, %f1094;
|
|
sub.f32 %f298, %f297, %f296;
|
|
mul.f32 %f299, %f1089, %f1094;
|
|
mul.f32 %f300, %f1091, %f1092;
|
|
sub.f32 %f301, %f300, %f299;
|
|
mul.f32 %f302, %f1090, %f1092;
|
|
mul.f32 %f303, %f1089, %f1093;
|
|
sub.f32 %f304, %f303, %f302;
|
|
setp.lt.u32 %p14, %r5, 16777216;
|
|
selp.f32 %f305, 0fBF800000, 0f3F800000, %p14;
|
|
mul.f32 %f306, %f298, %f305;
|
|
mul.f32 %f307, %f301, %f305;
|
|
mul.f32 %f308, %f304, %f305;
|
|
mul.f32 %f309, %f306, 0f00000000;
|
|
mul.f32 %f310, %f307, 0f00000000;
|
|
mul.f32 %f311, %f308, 0f00000000;
|
|
fma.rn.f32 %f312, %f1092, 0f3F5105EC, %f309;
|
|
fma.rn.f32 %f313, %f1093, 0f3F5105EC, %f310;
|
|
fma.rn.f32 %f314, %f1094, 0f3F5105EC, %f311;
|
|
mul.f32 %f21, %f1089, 0f3F13CD3A;
|
|
add.f32 %f22, %f21, %f312;
|
|
mul.f32 %f23, %f1090, 0f3F13CD3A;
|
|
add.f32 %f24, %f23, %f313;
|
|
mul.f32 %f25, %f1091, 0f3F13CD3A;
|
|
add.f32 %f26, %f25, %f314;
|
|
ld.global.v2.u32 {%r136, %r137}, [pixelID];
|
|
cvt.u64.u32 %rd47, %r136;
|
|
cvt.u64.u32 %rd48, %r137;
|
|
// inline asm
|
|
call (%rd45), _rt_buffer_get_64, (%rd38, %r107, %r108, %rd47, %rd48, %rd29, %rd29);
|
|
// inline asm
|
|
ld.u32 %r7, [%rd45];
|
|
shr.u32 %r8, %r7, 16;
|
|
cvt.u16.u32 %rs22, %r8;
|
|
and.b16 %rs23, %rs22, 255;
|
|
cvt.u16.u32 %rs24, %r7;
|
|
or.b16 %rs25, %rs24, %rs23;
|
|
setp.eq.s16 %p15, %rs25, 0;
|
|
mov.f32 %f1095, %f50;
|
|
mov.f32 %f1096, %f50;
|
|
mov.f32 %f1097, %f50;
|
|
@%p15 bra BB0_7;
|
|
|
|
ld.u8 %rs26, [%rd45+1];
|
|
and.b16 %rs28, %rs24, 255;
|
|
cvt.rn.f32.u16 %f315, %rs28;
|
|
div.rn.f32 %f316, %f315, 0f437F0000;
|
|
fma.rn.f32 %f317, %f316, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f318, %rs26;
|
|
div.rn.f32 %f319, %f318, 0f437F0000;
|
|
fma.rn.f32 %f320, %f319, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f321, %rs23;
|
|
div.rn.f32 %f322, %f321, 0f437F0000;
|
|
fma.rn.f32 %f323, %f322, 0f40000000, 0fBF800000;
|
|
mul.f32 %f324, %f320, %f320;
|
|
fma.rn.f32 %f325, %f317, %f317, %f324;
|
|
fma.rn.f32 %f326, %f323, %f323, %f325;
|
|
sqrt.rn.f32 %f327, %f326;
|
|
rcp.rn.f32 %f328, %f327;
|
|
mul.f32 %f1095, %f317, %f328;
|
|
mul.f32 %f1096, %f320, %f328;
|
|
mul.f32 %f1097, %f323, %f328;
|
|
|
|
BB0_7:
|
|
mul.f32 %f332, %f1091, %f1096;
|
|
mul.f32 %f333, %f1090, %f1097;
|
|
sub.f32 %f334, %f333, %f332;
|
|
mul.f32 %f335, %f1089, %f1097;
|
|
mul.f32 %f336, %f1091, %f1095;
|
|
sub.f32 %f337, %f336, %f335;
|
|
mul.f32 %f338, %f1090, %f1095;
|
|
mul.f32 %f339, %f1089, %f1096;
|
|
sub.f32 %f340, %f339, %f338;
|
|
setp.lt.u32 %p16, %r7, 16777216;
|
|
selp.f32 %f341, 0fBF800000, 0f3F800000, %p16;
|
|
mul.f32 %f342, %f334, %f341;
|
|
mul.f32 %f343, %f337, %f341;
|
|
mul.f32 %f344, %f340, %f341;
|
|
mul.f32 %f345, %f342, 0f3F3504F3;
|
|
mul.f32 %f346, %f343, 0f3F3504F3;
|
|
mul.f32 %f347, %f344, 0f3F3504F3;
|
|
fma.rn.f32 %f348, %f1095, 0fBED105EC, %f345;
|
|
fma.rn.f32 %f349, %f1096, 0fBED105EC, %f346;
|
|
fma.rn.f32 %f350, %f1097, 0fBED105EC, %f347;
|
|
add.f32 %f33, %f21, %f348;
|
|
add.f32 %f34, %f23, %f349;
|
|
add.f32 %f35, %f25, %f350;
|
|
ld.global.v2.u32 {%r142, %r143}, [pixelID];
|
|
cvt.u64.u32 %rd54, %r142;
|
|
cvt.u64.u32 %rd55, %r143;
|
|
// inline asm
|
|
call (%rd52), _rt_buffer_get_64, (%rd38, %r107, %r108, %rd54, %rd55, %rd29, %rd29);
|
|
// inline asm
|
|
ld.u32 %r9, [%rd52];
|
|
shr.u32 %r10, %r9, 16;
|
|
cvt.u16.u32 %rs31, %r10;
|
|
and.b16 %rs32, %rs31, 255;
|
|
cvt.u16.u32 %rs33, %r9;
|
|
or.b16 %rs34, %rs33, %rs32;
|
|
setp.eq.s16 %p17, %rs34, 0;
|
|
mov.f32 %f1098, %f50;
|
|
mov.f32 %f1099, %f50;
|
|
mov.f32 %f1100, %f50;
|
|
@%p17 bra BB0_9;
|
|
|
|
ld.u8 %rs35, [%rd52+1];
|
|
and.b16 %rs37, %rs33, 255;
|
|
cvt.rn.f32.u16 %f351, %rs37;
|
|
div.rn.f32 %f352, %f351, 0f437F0000;
|
|
fma.rn.f32 %f353, %f352, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f354, %rs35;
|
|
div.rn.f32 %f355, %f354, 0f437F0000;
|
|
fma.rn.f32 %f356, %f355, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f357, %rs32;
|
|
div.rn.f32 %f358, %f357, 0f437F0000;
|
|
fma.rn.f32 %f359, %f358, 0f40000000, 0fBF800000;
|
|
mul.f32 %f360, %f356, %f356;
|
|
fma.rn.f32 %f361, %f353, %f353, %f360;
|
|
fma.rn.f32 %f362, %f359, %f359, %f361;
|
|
sqrt.rn.f32 %f363, %f362;
|
|
rcp.rn.f32 %f364, %f363;
|
|
mul.f32 %f1098, %f353, %f364;
|
|
mul.f32 %f1099, %f356, %f364;
|
|
mul.f32 %f1100, %f359, %f364;
|
|
|
|
BB0_9:
|
|
mul.f32 %f370, %f1091, %f1099;
|
|
mul.f32 %f371, %f1090, %f1100;
|
|
sub.f32 %f372, %f371, %f370;
|
|
mul.f32 %f373, %f1089, %f1100;
|
|
mul.f32 %f374, %f1091, %f1098;
|
|
sub.f32 %f375, %f374, %f373;
|
|
mul.f32 %f376, %f1090, %f1098;
|
|
mul.f32 %f377, %f1089, %f1099;
|
|
sub.f32 %f378, %f377, %f376;
|
|
setp.lt.u32 %p19, %r9, 16777216;
|
|
selp.f32 %f379, 0fBF800000, 0f3F800000, %p19;
|
|
mul.f32 %f380, %f372, %f379;
|
|
mul.f32 %f381, %f375, %f379;
|
|
mul.f32 %f382, %f378, %f379;
|
|
mul.f32 %f383, %f380, 0fBF3504F3;
|
|
mul.f32 %f384, %f381, 0fBF3504F3;
|
|
mul.f32 %f385, %f382, 0fBF3504F3;
|
|
fma.rn.f32 %f386, %f1098, 0fBED105EC, %f383;
|
|
fma.rn.f32 %f387, %f1099, 0fBED105EC, %f384;
|
|
fma.rn.f32 %f388, %f1100, 0fBED105EC, %f385;
|
|
add.f32 %f42, %f21, %f386;
|
|
add.f32 %f43, %f23, %f387;
|
|
add.f32 %f44, %f25, %f388;
|
|
ld.global.v2.u32 {%r149, %r150}, [pixelID];
|
|
cvt.u64.u32 %rd61, %r149;
|
|
cvt.u64.u32 %rd62, %r150;
|
|
mov.u64 %rd65, rnd_seeds;
|
|
cvta.global.u64 %rd60, %rd65;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd60, %r107, %r108, %rd61, %rd62, %rd29, %rd29);
|
|
// inline asm
|
|
ld.u32 %r153, [%rd59];
|
|
mad.lo.s32 %r11, %r153, 1664525, 1013904223;
|
|
ld.global.u32 %r154, [lightMeshBufferSize];
|
|
setp.eq.s32 %p20, %r154, 0;
|
|
mov.pred %p18, 0;
|
|
mov.u32 %r13, 0;
|
|
@%p20 bra BB0_10;
|
|
|
|
ld.global.f32 %f45, [lightPointSize];
|
|
mul.f32 %f46, %f12, 0f3456BF95;
|
|
mul.f32 %f47, %f13, 0f3456BF95;
|
|
mul.f32 %f48, %f14, 0f3456BF95;
|
|
and.b32 %r157, %r11, 16777215;
|
|
cvt.rn.f32.u32 %f394, %r157;
|
|
mul.f32 %f395, %f394, 0fB3800000;
|
|
fma.rn.f32 %f49, %f395, 0f3F333333, 0f3F800000;
|
|
mov.f32 %f50, 0f00000000;
|
|
mov.u32 %r403, 0;
|
|
abs.f32 %f517, %f47;
|
|
abs.f32 %f518, %f46;
|
|
max.f32 %f519, %f518, %f517;
|
|
abs.f32 %f520, %f48;
|
|
max.f32 %f521, %f519, %f520;
|
|
mov.u32 %r13, %r403;
|
|
mov.f32 %f51, %f50;
|
|
mov.f32 %f52, %f50;
|
|
mov.f32 %f53, %f50;
|
|
mov.f32 %f54, %f50;
|
|
|
|
BB0_12:
|
|
shl.b32 %r14, %r403, 1;
|
|
cvt.s64.s32 %rd68, %r14;
|
|
mov.u64 %rd72, lightMeshBuffer;
|
|
cvta.global.u64 %rd67, %rd72;
|
|
mov.u32 %r158, 1;
|
|
// inline asm
|
|
call (%rd66), _rt_buffer_get_64, (%rd67, %r158, %r123, %rd68, %rd29, %rd29, %rd29);
|
|
// inline asm
|
|
ld.f32 %f396, [%rd66];
|
|
sub.f32 %f397, %f396, %f7;
|
|
ld.f32 %f398, [%rd66+4];
|
|
sub.f32 %f399, %f398, %f8;
|
|
ld.f32 %f400, [%rd66+8];
|
|
sub.f32 %f401, %f400, %f9;
|
|
mul.f32 %f402, %f399, %f399;
|
|
fma.rn.f32 %f403, %f397, %f397, %f402;
|
|
fma.rn.f32 %f404, %f401, %f401, %f403;
|
|
sqrt.rn.f32 %f55, %f404;
|
|
rcp.rn.f32 %f405, %f55;
|
|
mul.f32 %f56, %f397, %f405;
|
|
mul.f32 %f57, %f399, %f405;
|
|
mul.f32 %f58, %f401, %f405;
|
|
mul.f32 %f406, %f1090, %f57;
|
|
fma.rn.f32 %f407, %f1089, %f56, %f406;
|
|
fma.rn.f32 %f59, %f1091, %f58, %f407;
|
|
setp.leu.f32 %p21, %f59, 0f00000000;
|
|
@%p21 bra BB0_28;
|
|
|
|
setp.ne.s32 %p23, %r1, 0;
|
|
mul.f32 %f408, %f55, %f55;
|
|
mul.f32 %f409, %f408, 0f40C90FDB;
|
|
div.rn.f32 %f410, %f45, %f409;
|
|
add.f32 %f60, %f410, %f410;
|
|
setp.gt.f32 %p24, %f60, %f49;
|
|
and.pred %p25, %p23, %p24;
|
|
mov.pred %p155, -1;
|
|
@%p25 bra BB0_30;
|
|
|
|
ld.global.f32 %f413, [lightInvCutoff];
|
|
mul.f32 %f61, %f55, %f413;
|
|
mov.f32 %f417, 0f40800000;
|
|
abs.f32 %f63, %f61;
|
|
setp.lt.f32 %p26, %f63, 0f00800000;
|
|
mul.f32 %f419, %f63, 0f4B800000;
|
|
selp.f32 %f420, 0fC3170000, 0fC2FE0000, %p26;
|
|
selp.f32 %f421, %f419, %f63, %p26;
|
|
mov.b32 %r160, %f421;
|
|
and.b32 %r161, %r160, 8388607;
|
|
or.b32 %r162, %r161, 1065353216;
|
|
mov.b32 %f422, %r162;
|
|
shr.u32 %r163, %r160, 23;
|
|
cvt.rn.f32.u32 %f423, %r163;
|
|
add.f32 %f424, %f420, %f423;
|
|
setp.gt.f32 %p27, %f422, 0f3FB504F3;
|
|
mul.f32 %f425, %f422, 0f3F000000;
|
|
add.f32 %f426, %f424, 0f3F800000;
|
|
selp.f32 %f427, %f425, %f422, %p27;
|
|
selp.f32 %f428, %f426, %f424, %p27;
|
|
add.f32 %f429, %f427, 0fBF800000;
|
|
add.f32 %f412, %f427, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f411,%f412;
|
|
// inline asm
|
|
add.f32 %f430, %f429, %f429;
|
|
mul.f32 %f431, %f411, %f430;
|
|
mul.f32 %f432, %f431, %f431;
|
|
mov.f32 %f433, 0f3C4CAF63;
|
|
mov.f32 %f434, 0f3B18F0FE;
|
|
fma.rn.f32 %f435, %f434, %f432, %f433;
|
|
mov.f32 %f436, 0f3DAAAABD;
|
|
fma.rn.f32 %f437, %f435, %f432, %f436;
|
|
mul.rn.f32 %f438, %f437, %f432;
|
|
mul.rn.f32 %f439, %f438, %f431;
|
|
sub.f32 %f440, %f429, %f431;
|
|
neg.f32 %f441, %f431;
|
|
add.f32 %f442, %f440, %f440;
|
|
fma.rn.f32 %f443, %f441, %f429, %f442;
|
|
mul.rn.f32 %f444, %f411, %f443;
|
|
add.f32 %f445, %f439, %f431;
|
|
sub.f32 %f446, %f431, %f445;
|
|
add.f32 %f447, %f439, %f446;
|
|
add.f32 %f448, %f444, %f447;
|
|
add.f32 %f449, %f445, %f448;
|
|
sub.f32 %f450, %f445, %f449;
|
|
add.f32 %f451, %f448, %f450;
|
|
mov.f32 %f452, 0f3F317200;
|
|
mul.rn.f32 %f453, %f428, %f452;
|
|
mov.f32 %f454, 0f35BFBE8E;
|
|
mul.rn.f32 %f455, %f428, %f454;
|
|
add.f32 %f456, %f453, %f449;
|
|
sub.f32 %f457, %f453, %f456;
|
|
add.f32 %f458, %f449, %f457;
|
|
add.f32 %f459, %f451, %f458;
|
|
add.f32 %f460, %f455, %f459;
|
|
add.f32 %f461, %f456, %f460;
|
|
sub.f32 %f462, %f456, %f461;
|
|
add.f32 %f463, %f460, %f462;
|
|
mul.rn.f32 %f464, %f417, %f461;
|
|
neg.f32 %f465, %f464;
|
|
fma.rn.f32 %f466, %f417, %f461, %f465;
|
|
fma.rn.f32 %f467, %f417, %f463, %f466;
|
|
mov.f32 %f468, 0f00000000;
|
|
fma.rn.f32 %f469, %f468, %f461, %f467;
|
|
add.rn.f32 %f470, %f464, %f469;
|
|
neg.f32 %f471, %f470;
|
|
add.rn.f32 %f472, %f464, %f471;
|
|
add.rn.f32 %f473, %f472, %f469;
|
|
mov.b32 %r164, %f470;
|
|
setp.eq.s32 %p28, %r164, 1118925336;
|
|
add.s32 %r165, %r164, -1;
|
|
mov.b32 %f474, %r165;
|
|
add.f32 %f475, %f473, 0f37000000;
|
|
selp.f32 %f476, %f474, %f470, %p28;
|
|
selp.f32 %f64, %f475, %f473, %p28;
|
|
mul.f32 %f477, %f476, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f478, %f477;
|
|
mov.f32 %f479, 0fBF317200;
|
|
fma.rn.f32 %f480, %f478, %f479, %f476;
|
|
mov.f32 %f481, 0fB5BFBE8E;
|
|
fma.rn.f32 %f482, %f478, %f481, %f480;
|
|
mul.f32 %f483, %f482, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f484, %f483;
|
|
add.f32 %f485, %f478, 0f00000000;
|
|
ex2.approx.f32 %f486, %f485;
|
|
mul.f32 %f487, %f484, %f486;
|
|
setp.lt.f32 %p29, %f476, 0fC2D20000;
|
|
selp.f32 %f488, 0f00000000, %f487, %p29;
|
|
setp.gt.f32 %p30, %f476, 0f42D20000;
|
|
selp.f32 %f1106, 0f7F800000, %f488, %p30;
|
|
setp.eq.f32 %p31, %f1106, 0f7F800000;
|
|
@%p31 bra BB0_16;
|
|
|
|
fma.rn.f32 %f1106, %f1106, %f64, %f1106;
|
|
|
|
BB0_16:
|
|
mov.f32 %f1085, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1084, %f1085;
|
|
add.f32 %f1083, %f1084, %f1084;
|
|
mov.f32 %f1082, 0f40800000;
|
|
sub.f32 %f1081, %f1082, %f1083;
|
|
abs.f32 %f1080, %f1081;
|
|
setp.lt.f32 %p32, %f61, 0f00000000;
|
|
setp.eq.f32 %p33, %f1080, 0f3F800000;
|
|
and.pred %p1, %p32, %p33;
|
|
mov.b32 %r166, %f1106;
|
|
xor.b32 %r167, %r166, -2147483648;
|
|
mov.b32 %f489, %r167;
|
|
selp.f32 %f1108, %f489, %f1106, %p1;
|
|
setp.eq.f32 %p34, %f61, 0f00000000;
|
|
@%p34 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f492, %f61, %f61;
|
|
selp.f32 %f1108, %f492, 0f00000000, %p33;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.geu.f32 %p35, %f61, 0f00000000;
|
|
@%p35 bra BB0_20;
|
|
|
|
mov.f32 %f1088, 0f40800000;
|
|
cvt.rzi.f32.f32 %f491, %f1088;
|
|
setp.neu.f32 %p36, %f491, 0f40800000;
|
|
selp.f32 %f1108, 0f7FFFFFFF, %f1108, %p36;
|
|
|
|
BB0_20:
|
|
add.f32 %f493, %f63, 0f40800000;
|
|
mov.b32 %r168, %f493;
|
|
setp.lt.s32 %p38, %r168, 2139095040;
|
|
@%p38 bra BB0_25;
|
|
|
|
setp.gtu.f32 %p39, %f63, 0f7F800000;
|
|
@%p39 bra BB0_24;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_24:
|
|
add.f32 %f1108, %f61, 0f40800000;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_22:
|
|
setp.neu.f32 %p40, %f63, 0f7F800000;
|
|
@%p40 bra BB0_25;
|
|
|
|
selp.f32 %f1108, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_25:
|
|
mov.u32 %r390, 1;
|
|
mov.u64 %rd286, lightMeshBuffer;
|
|
cvta.global.u64 %rd285, %rd286;
|
|
shl.b32 %r389, %r403, 1;
|
|
mov.f32 %f494, 0f3F800000;
|
|
sub.f32 %f495, %f494, %f1108;
|
|
setp.eq.f32 %p41, %f61, 0f3F800000;
|
|
selp.f32 %f496, 0f00000000, %f495, %p41;
|
|
cvt.sat.f32.f32 %f497, %f496;
|
|
mul.f32 %f498, %f60, %f497;
|
|
add.s32 %r171, %r389, 1;
|
|
cvt.s64.s32 %rd75, %r171;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd285, %r390, %r123, %rd75, %rd29, %rd29, %rd29);
|
|
// inline asm
|
|
ld.f32 %f499, [%rd73];
|
|
mul.f32 %f500, %f56, %f499;
|
|
ld.f32 %f501, [%rd73+4];
|
|
mul.f32 %f502, %f57, %f501;
|
|
neg.f32 %f503, %f502;
|
|
sub.f32 %f504, %f503, %f500;
|
|
ld.f32 %f505, [%rd73+8];
|
|
mul.f32 %f506, %f58, %f505;
|
|
sub.f32 %f507, %f504, %f506;
|
|
cvt.sat.f32.f32 %f508, %f507;
|
|
mul.f32 %f75, %f498, %f508;
|
|
setp.leu.f32 %p42, %f75, 0f3727C5AC;
|
|
@%p42 bra BB0_27;
|
|
|
|
mov.u32 %r402, 1;
|
|
add.u64 %rd80, %SP, 4;
|
|
cvta.to.local.u64 %rd81, %rd80;
|
|
max.f32 %f515, %f521, %f275;
|
|
sub.f32 %f516, %f55, %f515;
|
|
mov.u32 %r175, 1065353216;
|
|
st.local.u32 [%rd81], %r175;
|
|
ld.global.u32 %r172, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r172, %f12, %f13, %f14, %f56, %f57, %f58, %r402, %f515, %f516, %rd80, %r108);
|
|
// inline asm
|
|
ld.local.f32 %f523, [%rd81];
|
|
mul.f32 %f524, %f75, %f523;
|
|
cvt.sat.f32.f32 %f525, %f59;
|
|
fma.rn.f32 %f54, %f525, %f524, %f54;
|
|
mul.f32 %f526, %f24, %f57;
|
|
fma.rn.f32 %f527, %f22, %f56, %f526;
|
|
fma.rn.f32 %f528, %f26, %f58, %f527;
|
|
cvt.sat.f32.f32 %f529, %f528;
|
|
fma.rn.f32 %f53, %f529, %f524, %f53;
|
|
mul.f32 %f530, %f34, %f57;
|
|
fma.rn.f32 %f531, %f33, %f56, %f530;
|
|
fma.rn.f32 %f532, %f35, %f58, %f531;
|
|
cvt.sat.f32.f32 %f533, %f532;
|
|
fma.rn.f32 %f52, %f524, %f533, %f52;
|
|
mul.f32 %f534, %f43, %f57;
|
|
fma.rn.f32 %f535, %f42, %f56, %f534;
|
|
fma.rn.f32 %f536, %f44, %f58, %f535;
|
|
cvt.sat.f32.f32 %f537, %f536;
|
|
fma.rn.f32 %f51, %f524, %f537, %f51;
|
|
add.f32 %f50, %f50, %f523;
|
|
|
|
BB0_27:
|
|
add.s32 %r13, %r13, 1;
|
|
|
|
BB0_28:
|
|
ld.global.u32 %r176, [lightMeshBufferSize];
|
|
add.s32 %r403, %r403, 1;
|
|
setp.lt.u32 %p44, %r403, %r176;
|
|
@%p44 bra BB0_12;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_10:
|
|
mov.f32 %f51, %f50;
|
|
mov.f32 %f52, %f50;
|
|
mov.f32 %f53, %f50;
|
|
mov.f32 %f54, %f50;
|
|
|
|
BB0_29:
|
|
mov.pred %p155, %p18;
|
|
|
|
BB0_30:
|
|
cvt.rn.f32.s32 %f538, %r13;
|
|
mov.f32 %f539, 0f3F800000;
|
|
max.f32 %f540, %f538, %f539;
|
|
div.rn.f32 %f1151, %f54, %f540;
|
|
div.rn.f32 %f1155, %f50, %f540;
|
|
div.rn.f32 %f1152, %f53, %f540;
|
|
div.rn.f32 %f1153, %f52, %f540;
|
|
div.rn.f32 %f1154, %f51, %f540;
|
|
@!%p155 bra BB0_83;
|
|
bra.uni BB0_31;
|
|
|
|
BB0_31:
|
|
abs.f32 %f1087, %f1091;
|
|
abs.f32 %f1086, %f1089;
|
|
setp.gt.f32 %p45, %f1086, %f1087;
|
|
neg.f32 %f546, %f1090;
|
|
selp.f32 %f547, %f546, 0f00000000, %p45;
|
|
neg.f32 %f548, %f1091;
|
|
selp.f32 %f549, %f1089, %f548, %p45;
|
|
selp.f32 %f550, 0f00000000, %f1090, %p45;
|
|
mul.f32 %f551, %f549, %f549;
|
|
fma.rn.f32 %f552, %f547, %f547, %f551;
|
|
fma.rn.f32 %f553, %f550, %f550, %f552;
|
|
sqrt.rn.f32 %f554, %f553;
|
|
rcp.rn.f32 %f555, %f554;
|
|
mul.f32 %f101, %f547, %f555;
|
|
mul.f32 %f102, %f549, %f555;
|
|
mul.f32 %f103, %f550, %f555;
|
|
mov.f32 %f1129, 0f00000000;
|
|
setp.lt.s32 %p46, %r1, 1;
|
|
mov.f32 %f1130, %f1129;
|
|
mov.f32 %f1131, %f1129;
|
|
mov.f32 %f1132, %f1129;
|
|
mov.f32 %f1133, %f1129;
|
|
@%p46 bra BB0_82;
|
|
|
|
mad.lo.s32 %r410, %r153, 1664525, 1013904223;
|
|
cvt.rn.f32.s32 %f561, %r1;
|
|
rcp.rn.f32 %f104, %f561;
|
|
add.u64 %rd82, %SP, 8;
|
|
cvta.to.local.u64 %rd5, %rd82;
|
|
mul.f32 %f105, %f12, 0f3456BF95;
|
|
mul.f32 %f106, %f13, 0f3456BF95;
|
|
mul.f32 %f107, %f14, 0f3456BF95;
|
|
add.u64 %rd83, %SP, 0;
|
|
cvta.to.local.u64 %rd6, %rd83;
|
|
mul.f32 %f562, %f1089, %f102;
|
|
mul.f32 %f563, %f1090, %f101;
|
|
sub.f32 %f108, %f563, %f562;
|
|
mul.f32 %f564, %f1091, %f101;
|
|
mul.f32 %f565, %f1089, %f103;
|
|
sub.f32 %f109, %f565, %f564;
|
|
mul.f32 %f566, %f1090, %f103;
|
|
mul.f32 %f567, %f1091, %f102;
|
|
sub.f32 %f110, %f567, %f566;
|
|
mov.f32 %f1129, 0f00000000;
|
|
mov.u32 %r177, 0;
|
|
abs.f32 %f568, %f106;
|
|
abs.f32 %f569, %f105;
|
|
max.f32 %f570, %f569, %f568;
|
|
abs.f32 %f571, %f107;
|
|
max.f32 %f572, %f570, %f571;
|
|
mov.u32 %r407, %r177;
|
|
mov.f32 %f1130, %f1129;
|
|
mov.f32 %f1131, %f1129;
|
|
mov.f32 %f1132, %f1129;
|
|
mov.f32 %f1133, %f1129;
|
|
|
|
BB0_33:
|
|
cvt.rn.f32.s32 %f116, %r407;
|
|
max.f32 %f117, %f572, %f275;
|
|
mov.u32 %r409, %r177;
|
|
|
|
BB0_34:
|
|
mad.lo.s32 %r179, %r410, 1664525, 1013904223;
|
|
and.b32 %r180, %r179, 16777215;
|
|
cvt.rn.f32.u32 %f574, %r180;
|
|
fma.rn.f32 %f575, %f574, 0f33800000, %f116;
|
|
mul.f32 %f123, %f104, %f575;
|
|
mad.lo.s32 %r410, %r179, 1664525, 1013904223;
|
|
and.b32 %r181, %r410, 16777215;
|
|
cvt.rn.f32.u32 %f576, %r181;
|
|
cvt.rn.f32.s32 %f577, %r409;
|
|
fma.rn.f32 %f578, %f576, 0f33800000, %f577;
|
|
mul.f32 %f579, %f104, %f578;
|
|
mul.f32 %f580, %f123, %f123;
|
|
sub.f32 %f582, %f539, %f580;
|
|
mov.f32 %f583, 0f00000000;
|
|
max.f32 %f584, %f583, %f582;
|
|
sqrt.rn.f32 %f124, %f584;
|
|
mul.f32 %f1140, %f579, 0f40C90FDB;
|
|
abs.f32 %f126, %f1140;
|
|
setp.neu.f32 %p47, %f126, 0f7F800000;
|
|
mov.f32 %f1134, %f1140;
|
|
@%p47 bra BB0_36;
|
|
|
|
mul.rn.f32 %f1134, %f1140, %f583;
|
|
|
|
BB0_36:
|
|
mul.f32 %f586, %f1134, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r420, %f586;
|
|
cvt.rn.f32.s32 %f587, %r420;
|
|
neg.f32 %f588, %f587;
|
|
mov.f32 %f589, 0f3FC90FDA;
|
|
fma.rn.f32 %f590, %f588, %f589, %f1134;
|
|
mov.f32 %f591, 0f33A22168;
|
|
fma.rn.f32 %f592, %f588, %f591, %f590;
|
|
mov.f32 %f593, 0f27C234C5;
|
|
fma.rn.f32 %f1135, %f588, %f593, %f592;
|
|
abs.f32 %f594, %f1134;
|
|
setp.leu.f32 %p48, %f594, 0f47CE4780;
|
|
@%p48 bra BB0_47;
|
|
|
|
mov.b32 %r25, %f1134;
|
|
shr.u32 %r26, %r25, 23;
|
|
shl.b32 %r184, %r25, 8;
|
|
or.b32 %r27, %r184, -2147483648;
|
|
mov.u32 %r411, 0;
|
|
mov.u64 %rd288, 0;
|
|
mov.u64 %rd287, %rd5;
|
|
mov.u32 %r412, %r411;
|
|
|
|
BB0_38:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd85, %rd288, 2;
|
|
mov.u64 %rd86, __cudart_i2opi_f;
|
|
add.s64 %rd87, %rd86, %rd85;
|
|
ld.const.u32 %r187, [%rd87];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r185, %r187, %r27, %r412;
|
|
madc.hi.u32 %r412, %r187, %r27, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd287], %r185;
|
|
add.s32 %r411, %r411, 1;
|
|
cvt.s64.s32 %rd288, %r411;
|
|
mul.wide.s32 %rd88, %r411, 4;
|
|
add.s64 %rd287, %rd5, %rd88;
|
|
setp.ne.s32 %p49, %r411, 6;
|
|
@%p49 bra BB0_38;
|
|
|
|
and.b32 %r190, %r26, 255;
|
|
add.s32 %r191, %r190, -128;
|
|
shr.u32 %r192, %r191, 5;
|
|
and.b32 %r32, %r25, -2147483648;
|
|
st.local.u32 [%rd5+24], %r412;
|
|
mov.u32 %r193, 6;
|
|
sub.s32 %r194, %r193, %r192;
|
|
mul.wide.s32 %rd89, %r194, 4;
|
|
add.s64 %rd12, %rd5, %rd89;
|
|
ld.local.u32 %r413, [%rd12];
|
|
ld.local.u32 %r414, [%rd12+-4];
|
|
and.b32 %r35, %r26, 31;
|
|
setp.eq.s32 %p50, %r35, 0;
|
|
@%p50 bra BB0_41;
|
|
|
|
mov.u32 %r195, 32;
|
|
sub.s32 %r196, %r195, %r35;
|
|
shr.u32 %r197, %r414, %r196;
|
|
shl.b32 %r198, %r413, %r35;
|
|
add.s32 %r413, %r197, %r198;
|
|
ld.local.u32 %r199, [%rd12+-8];
|
|
shr.u32 %r200, %r199, %r196;
|
|
shl.b32 %r201, %r414, %r35;
|
|
add.s32 %r414, %r200, %r201;
|
|
|
|
BB0_41:
|
|
shr.u32 %r202, %r414, 30;
|
|
shl.b32 %r203, %r413, 2;
|
|
add.s32 %r415, %r202, %r203;
|
|
shl.b32 %r41, %r414, 2;
|
|
shr.u32 %r204, %r415, 31;
|
|
shr.u32 %r205, %r413, 30;
|
|
add.s32 %r42, %r204, %r205;
|
|
setp.eq.s32 %p51, %r204, 0;
|
|
@%p51 bra BB0_42;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_42:
|
|
mov.u32 %r416, %r32;
|
|
mov.u32 %r417, %r41;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_43:
|
|
not.b32 %r206, %r415;
|
|
neg.s32 %r417, %r41;
|
|
setp.eq.s32 %p52, %r41, 0;
|
|
selp.u32 %r207, 1, 0, %p52;
|
|
add.s32 %r415, %r207, %r206;
|
|
xor.b32 %r416, %r32, -2147483648;
|
|
|
|
BB0_44:
|
|
clz.b32 %r419, %r415;
|
|
setp.eq.s32 %p53, %r419, 0;
|
|
shl.b32 %r208, %r415, %r419;
|
|
mov.u32 %r209, 32;
|
|
sub.s32 %r210, %r209, %r419;
|
|
shr.u32 %r211, %r417, %r210;
|
|
add.s32 %r212, %r211, %r208;
|
|
selp.b32 %r50, %r415, %r212, %p53;
|
|
mov.u32 %r213, -921707870;
|
|
mul.hi.u32 %r418, %r50, %r213;
|
|
setp.eq.s32 %p54, %r32, 0;
|
|
neg.s32 %r214, %r42;
|
|
selp.b32 %r420, %r42, %r214, %p54;
|
|
setp.lt.s32 %p55, %r418, 1;
|
|
@%p55 bra BB0_46;
|
|
|
|
mul.lo.s32 %r215, %r50, -921707870;
|
|
shr.u32 %r216, %r215, 31;
|
|
shl.b32 %r217, %r418, 1;
|
|
add.s32 %r418, %r216, %r217;
|
|
add.s32 %r419, %r419, 1;
|
|
|
|
BB0_46:
|
|
mov.u32 %r218, 126;
|
|
sub.s32 %r219, %r218, %r419;
|
|
shl.b32 %r220, %r219, 23;
|
|
add.s32 %r221, %r418, 1;
|
|
shr.u32 %r222, %r221, 7;
|
|
add.s32 %r223, %r222, 1;
|
|
shr.u32 %r224, %r223, 1;
|
|
add.s32 %r225, %r224, %r220;
|
|
or.b32 %r226, %r225, %r416;
|
|
mov.b32 %f1135, %r226;
|
|
|
|
BB0_47:
|
|
mul.rn.f32 %f132, %f1135, %f1135;
|
|
add.s32 %r58, %r420, 1;
|
|
and.b32 %r59, %r58, 1;
|
|
setp.eq.s32 %p56, %r59, 0;
|
|
@%p56 bra BB0_49;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_49:
|
|
mov.f32 %f597, 0f3C08839E;
|
|
mov.f32 %f598, 0fB94CA1F9;
|
|
fma.rn.f32 %f1136, %f598, %f132, %f597;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_48:
|
|
mov.f32 %f595, 0fBAB6061A;
|
|
mov.f32 %f596, 0f37CCF5CE;
|
|
fma.rn.f32 %f1136, %f596, %f132, %f595;
|
|
|
|
BB0_50:
|
|
@%p56 bra BB0_52;
|
|
bra.uni BB0_51;
|
|
|
|
BB0_52:
|
|
mov.f32 %f602, 0fBE2AAAA3;
|
|
fma.rn.f32 %f603, %f1136, %f132, %f602;
|
|
fma.rn.f32 %f1137, %f603, %f132, %f583;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_51:
|
|
mov.f32 %f599, 0f3D2AAAA5;
|
|
fma.rn.f32 %f600, %f1136, %f132, %f599;
|
|
mov.f32 %f601, 0fBF000000;
|
|
fma.rn.f32 %f1137, %f600, %f132, %f601;
|
|
|
|
BB0_53:
|
|
fma.rn.f32 %f1138, %f1137, %f1135, %f1135;
|
|
@%p56 bra BB0_55;
|
|
|
|
fma.rn.f32 %f1138, %f1137, %f132, %f539;
|
|
|
|
BB0_55:
|
|
and.b32 %r227, %r58, 2;
|
|
setp.eq.s32 %p59, %r227, 0;
|
|
@%p59 bra BB0_57;
|
|
|
|
mov.f32 %f607, 0fBF800000;
|
|
fma.rn.f32 %f1138, %f1138, %f607, %f583;
|
|
|
|
BB0_57:
|
|
@%p47 bra BB0_59;
|
|
|
|
mul.rn.f32 %f1140, %f1140, %f583;
|
|
|
|
BB0_59:
|
|
mul.f32 %f609, %f1140, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r430, %f609;
|
|
cvt.rn.f32.s32 %f610, %r430;
|
|
neg.f32 %f611, %f610;
|
|
fma.rn.f32 %f613, %f611, %f589, %f1140;
|
|
fma.rn.f32 %f615, %f611, %f591, %f613;
|
|
fma.rn.f32 %f1141, %f611, %f593, %f615;
|
|
abs.f32 %f617, %f1140;
|
|
setp.leu.f32 %p61, %f617, 0f47CE4780;
|
|
@%p61 bra BB0_70;
|
|
|
|
mov.b32 %r61, %f1140;
|
|
shr.u32 %r62, %r61, 23;
|
|
shl.b32 %r230, %r61, 8;
|
|
or.b32 %r63, %r230, -2147483648;
|
|
mov.u32 %r421, 0;
|
|
mov.u64 %rd289, %rd5;
|
|
mov.u64 %rd290, %rd29;
|
|
mov.u32 %r422, %r421;
|
|
|
|
BB0_61:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd91, %rd290, 2;
|
|
mov.u64 %rd92, __cudart_i2opi_f;
|
|
add.s64 %rd93, %rd92, %rd91;
|
|
ld.const.u32 %r233, [%rd93];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r231, %r233, %r63, %r422;
|
|
madc.hi.u32 %r422, %r233, %r63, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd289], %r231;
|
|
add.s32 %r421, %r421, 1;
|
|
cvt.s64.s32 %rd290, %r421;
|
|
mul.wide.s32 %rd94, %r421, 4;
|
|
add.s64 %rd289, %rd5, %rd94;
|
|
setp.ne.s32 %p62, %r421, 6;
|
|
@%p62 bra BB0_61;
|
|
|
|
and.b32 %r236, %r62, 255;
|
|
add.s32 %r237, %r236, -128;
|
|
shr.u32 %r238, %r237, 5;
|
|
and.b32 %r68, %r61, -2147483648;
|
|
st.local.u32 [%rd5+24], %r422;
|
|
mov.u32 %r239, 6;
|
|
sub.s32 %r240, %r239, %r238;
|
|
mul.wide.s32 %rd95, %r240, 4;
|
|
add.s64 %rd18, %rd5, %rd95;
|
|
ld.local.u32 %r423, [%rd18];
|
|
ld.local.u32 %r424, [%rd18+-4];
|
|
and.b32 %r71, %r62, 31;
|
|
setp.eq.s32 %p63, %r71, 0;
|
|
@%p63 bra BB0_64;
|
|
|
|
mov.u32 %r241, 32;
|
|
sub.s32 %r242, %r241, %r71;
|
|
shr.u32 %r243, %r424, %r242;
|
|
shl.b32 %r244, %r423, %r71;
|
|
add.s32 %r423, %r243, %r244;
|
|
ld.local.u32 %r245, [%rd18+-8];
|
|
shr.u32 %r246, %r245, %r242;
|
|
shl.b32 %r247, %r424, %r71;
|
|
add.s32 %r424, %r246, %r247;
|
|
|
|
BB0_64:
|
|
shr.u32 %r248, %r424, 30;
|
|
shl.b32 %r249, %r423, 2;
|
|
add.s32 %r425, %r248, %r249;
|
|
shl.b32 %r77, %r424, 2;
|
|
shr.u32 %r250, %r425, 31;
|
|
shr.u32 %r251, %r423, 30;
|
|
add.s32 %r78, %r250, %r251;
|
|
setp.eq.s32 %p64, %r250, 0;
|
|
@%p64 bra BB0_65;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_65:
|
|
mov.u32 %r426, %r68;
|
|
mov.u32 %r427, %r77;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_66:
|
|
not.b32 %r252, %r425;
|
|
neg.s32 %r427, %r77;
|
|
setp.eq.s32 %p65, %r77, 0;
|
|
selp.u32 %r253, 1, 0, %p65;
|
|
add.s32 %r425, %r253, %r252;
|
|
xor.b32 %r426, %r68, -2147483648;
|
|
|
|
BB0_67:
|
|
clz.b32 %r429, %r425;
|
|
setp.eq.s32 %p66, %r429, 0;
|
|
shl.b32 %r254, %r425, %r429;
|
|
mov.u32 %r255, 32;
|
|
sub.s32 %r256, %r255, %r429;
|
|
shr.u32 %r257, %r427, %r256;
|
|
add.s32 %r258, %r257, %r254;
|
|
selp.b32 %r86, %r425, %r258, %p66;
|
|
mov.u32 %r259, -921707870;
|
|
mul.hi.u32 %r428, %r86, %r259;
|
|
setp.eq.s32 %p67, %r68, 0;
|
|
neg.s32 %r260, %r78;
|
|
selp.b32 %r430, %r78, %r260, %p67;
|
|
setp.lt.s32 %p68, %r428, 1;
|
|
@%p68 bra BB0_69;
|
|
|
|
mul.lo.s32 %r261, %r86, -921707870;
|
|
shr.u32 %r262, %r261, 31;
|
|
shl.b32 %r263, %r428, 1;
|
|
add.s32 %r428, %r262, %r263;
|
|
add.s32 %r429, %r429, 1;
|
|
|
|
BB0_69:
|
|
mov.u32 %r264, 126;
|
|
sub.s32 %r265, %r264, %r429;
|
|
shl.b32 %r266, %r265, 23;
|
|
add.s32 %r267, %r428, 1;
|
|
shr.u32 %r268, %r267, 7;
|
|
add.s32 %r269, %r268, 1;
|
|
shr.u32 %r270, %r269, 1;
|
|
add.s32 %r271, %r270, %r266;
|
|
or.b32 %r272, %r271, %r426;
|
|
mov.b32 %f1141, %r272;
|
|
|
|
BB0_70:
|
|
mul.rn.f32 %f149, %f1141, %f1141;
|
|
and.b32 %r94, %r430, 1;
|
|
setp.eq.s32 %p69, %r94, 0;
|
|
@%p69 bra BB0_72;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_72:
|
|
mov.f32 %f620, 0f3C08839E;
|
|
mov.f32 %f621, 0fB94CA1F9;
|
|
fma.rn.f32 %f1142, %f621, %f149, %f620;
|
|
bra.uni BB0_73;
|
|
|
|
BB0_71:
|
|
mov.f32 %f618, 0fBAB6061A;
|
|
mov.f32 %f619, 0f37CCF5CE;
|
|
fma.rn.f32 %f1142, %f619, %f149, %f618;
|
|
|
|
BB0_73:
|
|
@%p69 bra BB0_75;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_75:
|
|
mov.f32 %f625, 0fBE2AAAA3;
|
|
fma.rn.f32 %f626, %f1142, %f149, %f625;
|
|
fma.rn.f32 %f1143, %f626, %f149, %f583;
|
|
bra.uni BB0_76;
|
|
|
|
BB0_74:
|
|
mov.f32 %f622, 0f3D2AAAA5;
|
|
fma.rn.f32 %f623, %f1142, %f149, %f622;
|
|
mov.f32 %f624, 0fBF000000;
|
|
fma.rn.f32 %f1143, %f623, %f149, %f624;
|
|
|
|
BB0_76:
|
|
fma.rn.f32 %f1144, %f1143, %f1141, %f1141;
|
|
@%p69 bra BB0_78;
|
|
|
|
fma.rn.f32 %f1144, %f1143, %f149, %f539;
|
|
|
|
BB0_78:
|
|
and.b32 %r273, %r430, 2;
|
|
setp.eq.s32 %p72, %r273, 0;
|
|
@%p72 bra BB0_80;
|
|
|
|
mov.f32 %f630, 0fBF800000;
|
|
fma.rn.f32 %f1144, %f1144, %f630, %f583;
|
|
|
|
BB0_80:
|
|
mul.f32 %f639, %f124, %f1138;
|
|
mul.f32 %f640, %f124, %f1144;
|
|
mul.f32 %f641, %f101, %f640;
|
|
mul.f32 %f642, %f102, %f640;
|
|
mul.f32 %f643, %f103, %f640;
|
|
fma.rn.f32 %f644, %f110, %f639, %f641;
|
|
fma.rn.f32 %f645, %f109, %f639, %f642;
|
|
fma.rn.f32 %f646, %f108, %f639, %f643;
|
|
fma.rn.f32 %f634, %f1089, %f123, %f644;
|
|
fma.rn.f32 %f635, %f1090, %f123, %f645;
|
|
fma.rn.f32 %f636, %f1091, %f123, %f646;
|
|
mov.u32 %r275, 0;
|
|
st.local.u32 [%rd6], %r275;
|
|
ld.global.u32 %r274, [root];
|
|
mov.f32 %f638, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r274, %f12, %f13, %f14, %f634, %f635, %f636, %r275, %f117, %f638, %rd83, %r108);
|
|
// inline asm
|
|
ld.local.f32 %f647, [%rd6];
|
|
setp.lt.f32 %p73, %f647, 0f00000000;
|
|
selp.f32 %f648, 0f00000000, %f647, %p73;
|
|
selp.f32 %f649, 0f00000000, 0f3F800000, %p73;
|
|
add.f32 %f1129, %f1129, %f649;
|
|
mul.f32 %f650, %f24, %f635;
|
|
fma.rn.f32 %f651, %f22, %f634, %f650;
|
|
fma.rn.f32 %f652, %f26, %f636, %f651;
|
|
cvt.sat.f32.f32 %f653, %f652;
|
|
fma.rn.f32 %f1132, %f648, %f653, %f1132;
|
|
mul.f32 %f654, %f34, %f635;
|
|
fma.rn.f32 %f655, %f33, %f634, %f654;
|
|
fma.rn.f32 %f656, %f35, %f636, %f655;
|
|
cvt.sat.f32.f32 %f657, %f656;
|
|
fma.rn.f32 %f1131, %f648, %f657, %f1131;
|
|
mul.f32 %f658, %f43, %f635;
|
|
fma.rn.f32 %f659, %f42, %f634, %f658;
|
|
fma.rn.f32 %f660, %f44, %f636, %f659;
|
|
cvt.sat.f32.f32 %f661, %f660;
|
|
fma.rn.f32 %f1130, %f648, %f661, %f1130;
|
|
mul.f32 %f662, %f1090, %f635;
|
|
fma.rn.f32 %f663, %f1089, %f634, %f662;
|
|
fma.rn.f32 %f664, %f1091, %f636, %f663;
|
|
cvt.sat.f32.f32 %f665, %f664;
|
|
fma.rn.f32 %f1133, %f648, %f665, %f1133;
|
|
add.s32 %r409, %r409, 1;
|
|
setp.lt.s32 %p74, %r409, %r1;
|
|
@%p74 bra BB0_34;
|
|
|
|
add.s32 %r407, %r407, 1;
|
|
setp.lt.s32 %p75, %r407, %r1;
|
|
@%p75 bra BB0_33;
|
|
|
|
BB0_82:
|
|
mul.lo.s32 %r277, %r1, %r1;
|
|
cvt.rn.f32.s32 %f666, %r277;
|
|
div.rn.f32 %f667, %f1133, %f666;
|
|
div.rn.f32 %f1155, %f1129, %f666;
|
|
div.rn.f32 %f1152, %f1132, %f666;
|
|
div.rn.f32 %f1153, %f1131, %f666;
|
|
div.rn.f32 %f1154, %f1130, %f666;
|
|
add.f32 %f1151, %f667, %f667;
|
|
|
|
BB0_83:
|
|
ld.global.u32 %r432, [imageEnabled];
|
|
and.b32 %r278, %r432, 8;
|
|
setp.eq.s32 %p76, %r278, 0;
|
|
@%p76 bra BB0_96;
|
|
|
|
mov.u32 %r391, 2;
|
|
cvt.u64.u32 %rd99, %r3;
|
|
cvt.u64.u32 %rd100, %r4;
|
|
mov.u64 %rd103, image_Mask;
|
|
cvta.global.u64 %rd98, %rd103;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r391, %r391, %rd99, %rd100, %rd29, %rd29);
|
|
// inline asm
|
|
mov.f32 %f670, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f671, %f670;
|
|
fma.rn.f32 %f672, %f671, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f181, %f672;
|
|
abs.f32 %f182, %f1155;
|
|
setp.lt.f32 %p77, %f182, 0f00800000;
|
|
mul.f32 %f673, %f182, 0f4B800000;
|
|
selp.f32 %f674, 0fC3170000, 0fC2FE0000, %p77;
|
|
selp.f32 %f675, %f673, %f182, %p77;
|
|
mov.b32 %r281, %f675;
|
|
and.b32 %r282, %r281, 8388607;
|
|
or.b32 %r283, %r282, 1065353216;
|
|
mov.b32 %f676, %r283;
|
|
shr.u32 %r284, %r281, 23;
|
|
cvt.rn.f32.u32 %f677, %r284;
|
|
add.f32 %f678, %f674, %f677;
|
|
setp.gt.f32 %p78, %f676, 0f3FB504F3;
|
|
mul.f32 %f679, %f676, 0f3F000000;
|
|
add.f32 %f680, %f678, 0f3F800000;
|
|
selp.f32 %f681, %f679, %f676, %p78;
|
|
selp.f32 %f682, %f680, %f678, %p78;
|
|
add.f32 %f683, %f681, 0fBF800000;
|
|
add.f32 %f669, %f681, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f668,%f669;
|
|
// inline asm
|
|
add.f32 %f684, %f683, %f683;
|
|
mul.f32 %f685, %f668, %f684;
|
|
mul.f32 %f686, %f685, %f685;
|
|
mov.f32 %f687, 0f3C4CAF63;
|
|
mov.f32 %f688, 0f3B18F0FE;
|
|
fma.rn.f32 %f689, %f688, %f686, %f687;
|
|
mov.f32 %f690, 0f3DAAAABD;
|
|
fma.rn.f32 %f691, %f689, %f686, %f690;
|
|
mul.rn.f32 %f692, %f691, %f686;
|
|
mul.rn.f32 %f693, %f692, %f685;
|
|
sub.f32 %f694, %f683, %f685;
|
|
neg.f32 %f695, %f685;
|
|
add.f32 %f696, %f694, %f694;
|
|
fma.rn.f32 %f697, %f695, %f683, %f696;
|
|
mul.rn.f32 %f698, %f668, %f697;
|
|
add.f32 %f699, %f693, %f685;
|
|
sub.f32 %f700, %f685, %f699;
|
|
add.f32 %f701, %f693, %f700;
|
|
add.f32 %f702, %f698, %f701;
|
|
add.f32 %f703, %f699, %f702;
|
|
sub.f32 %f704, %f699, %f703;
|
|
add.f32 %f705, %f702, %f704;
|
|
mov.f32 %f706, 0f3F317200;
|
|
mul.rn.f32 %f707, %f682, %f706;
|
|
mov.f32 %f708, 0f35BFBE8E;
|
|
mul.rn.f32 %f709, %f682, %f708;
|
|
add.f32 %f710, %f707, %f703;
|
|
sub.f32 %f711, %f707, %f710;
|
|
add.f32 %f712, %f703, %f711;
|
|
add.f32 %f713, %f705, %f712;
|
|
add.f32 %f714, %f709, %f713;
|
|
add.f32 %f715, %f710, %f714;
|
|
sub.f32 %f716, %f710, %f715;
|
|
add.f32 %f717, %f714, %f716;
|
|
mov.f32 %f718, 0f3EE8BA2E;
|
|
mul.rn.f32 %f719, %f718, %f715;
|
|
neg.f32 %f720, %f719;
|
|
fma.rn.f32 %f721, %f718, %f715, %f720;
|
|
fma.rn.f32 %f722, %f718, %f717, %f721;
|
|
mov.f32 %f723, 0f00000000;
|
|
fma.rn.f32 %f724, %f723, %f715, %f722;
|
|
add.rn.f32 %f725, %f719, %f724;
|
|
neg.f32 %f726, %f725;
|
|
add.rn.f32 %f727, %f719, %f726;
|
|
add.rn.f32 %f728, %f727, %f724;
|
|
mov.b32 %r285, %f725;
|
|
setp.eq.s32 %p79, %r285, 1118925336;
|
|
add.s32 %r286, %r285, -1;
|
|
mov.b32 %f729, %r286;
|
|
add.f32 %f730, %f728, 0f37000000;
|
|
selp.f32 %f731, %f729, %f725, %p79;
|
|
selp.f32 %f183, %f730, %f728, %p79;
|
|
mul.f32 %f732, %f731, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f733, %f732;
|
|
mov.f32 %f734, 0fBF317200;
|
|
fma.rn.f32 %f735, %f733, %f734, %f731;
|
|
mov.f32 %f736, 0fB5BFBE8E;
|
|
fma.rn.f32 %f737, %f733, %f736, %f735;
|
|
mul.f32 %f738, %f737, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f739, %f738;
|
|
add.f32 %f740, %f733, 0f00000000;
|
|
ex2.approx.f32 %f741, %f740;
|
|
mul.f32 %f742, %f739, %f741;
|
|
setp.lt.f32 %p80, %f731, 0fC2D20000;
|
|
selp.f32 %f743, 0f00000000, %f742, %p80;
|
|
setp.gt.f32 %p81, %f731, 0f42D20000;
|
|
selp.f32 %f1156, 0f7F800000, %f743, %p81;
|
|
setp.eq.f32 %p82, %f1156, 0f7F800000;
|
|
@%p82 bra BB0_86;
|
|
|
|
fma.rn.f32 %f1156, %f1156, %f183, %f1156;
|
|
|
|
BB0_86:
|
|
setp.lt.f32 %p83, %f1155, 0f00000000;
|
|
setp.eq.f32 %p84, %f181, 0f3F800000;
|
|
and.pred %p3, %p83, %p84;
|
|
mov.b32 %r287, %f1156;
|
|
xor.b32 %r288, %r287, -2147483648;
|
|
mov.b32 %f744, %r288;
|
|
selp.f32 %f1158, %f744, %f1156, %p3;
|
|
setp.eq.f32 %p85, %f1155, 0f00000000;
|
|
@%p85 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f747, %f1155, %f1155;
|
|
selp.f32 %f1158, %f747, 0f00000000, %p84;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_151:
|
|
mov.u64 %rd224, image_HDR;
|
|
cvta.global.u64 %rd219, %rd224;
|
|
mov.u32 %r367, 8;
|
|
mov.u64 %rd223, 0;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r107, %r367, %rd22, %rd23, %rd223, %rd223);
|
|
// inline asm
|
|
mov.f32 %f1058, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f1058;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs117, 0;
|
|
st.v4.u16 [%rd218], {%rs116, %rs116, %rs116, %rs117};
|
|
|
|
BB0_152:
|
|
ld.global.u32 %r368, [additive];
|
|
setp.eq.s32 %p152, %r368, 0;
|
|
@%p152 bra BB0_154;
|
|
|
|
mov.u64 %rd237, image_RNM0;
|
|
cvta.global.u64 %rd226, %rd237;
|
|
mov.u32 %r372, 8;
|
|
mov.u64 %rd236, 0;
|
|
// inline asm
|
|
call (%rd225), _rt_buffer_get_64, (%rd226, %r107, %r372, %rd22, %rd23, %rd236, %rd236);
|
|
// inline asm
|
|
ld.v4.u16 {%rs124, %rs125, %rs126, %rs127}, [%rd225];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1059, %rs124;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1060, %rs125;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1061, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd226, %r107, %r372, %rd22, %rd23, %rd236, %rd236);
|
|
// inline asm
|
|
add.f32 %f1062, %f1059, 0f00000000;
|
|
add.f32 %f1063, %f1060, 0f00000000;
|
|
add.f32 %f1064, %f1061, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1064;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f1063;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f1062;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs128, 0;
|
|
st.v4.u16 [%rd231], {%rs121, %rs122, %rs123, %rs128};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_154:
|
|
mov.u64 %rd244, image_RNM0;
|
|
cvta.global.u64 %rd239, %rd244;
|
|
mov.u32 %r374, 8;
|
|
mov.u64 %rd243, 0;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r107, %r374, %rd22, %rd23, %rd243, %rd243);
|
|
// inline asm
|
|
mov.f32 %f1065, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f1065;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs130, 0;
|
|
st.v4.u16 [%rd238], {%rs129, %rs129, %rs129, %rs130};
|
|
|
|
BB0_155:
|
|
ld.global.u32 %r375, [additive];
|
|
setp.eq.s32 %p153, %r375, 0;
|
|
@%p153 bra BB0_157;
|
|
|
|
mov.u64 %rd257, image_RNM1;
|
|
cvta.global.u64 %rd246, %rd257;
|
|
mov.u32 %r379, 8;
|
|
mov.u64 %rd256, 0;
|
|
// inline asm
|
|
call (%rd245), _rt_buffer_get_64, (%rd246, %r107, %r379, %rd22, %rd23, %rd256, %rd256);
|
|
// inline asm
|
|
ld.v4.u16 {%rs137, %rs138, %rs139, %rs140}, [%rd245];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1066, %rs137;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1067, %rs138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1068, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd246, %r107, %r379, %rd22, %rd23, %rd256, %rd256);
|
|
// inline asm
|
|
add.f32 %f1069, %f1066, 0f00000000;
|
|
add.f32 %f1070, %f1067, 0f00000000;
|
|
add.f32 %f1071, %f1068, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1071;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f1070;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f1069;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs141, 0;
|
|
st.v4.u16 [%rd251], {%rs134, %rs135, %rs136, %rs141};
|
|
bra.uni BB0_158;
|
|
|
|
BB0_157:
|
|
mov.u64 %rd264, image_RNM1;
|
|
cvta.global.u64 %rd259, %rd264;
|
|
mov.u32 %r381, 8;
|
|
mov.u64 %rd263, 0;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r107, %r381, %rd22, %rd23, %rd263, %rd263);
|
|
// inline asm
|
|
mov.f32 %f1072, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f1072;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs143, 0;
|
|
st.v4.u16 [%rd258], {%rs142, %rs142, %rs142, %rs143};
|
|
|
|
BB0_158:
|
|
ld.global.u32 %r382, [additive];
|
|
setp.eq.s32 %p154, %r382, 0;
|
|
@%p154 bra BB0_160;
|
|
|
|
mov.u64 %rd277, image_RNM2;
|
|
cvta.global.u64 %rd266, %rd277;
|
|
mov.u32 %r386, 8;
|
|
mov.u64 %rd276, 0;
|
|
// inline asm
|
|
call (%rd265), _rt_buffer_get_64, (%rd266, %r107, %r386, %rd22, %rd23, %rd276, %rd276);
|
|
// inline asm
|
|
ld.v4.u16 {%rs150, %rs151, %rs152, %rs153}, [%rd265];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1073, %rs150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1074, %rs151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1075, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd266, %r107, %r386, %rd22, %rd23, %rd276, %rd276);
|
|
// inline asm
|
|
add.f32 %f1076, %f1073, 0f00000000;
|
|
add.f32 %f1077, %f1074, 0f00000000;
|
|
add.f32 %f1078, %f1075, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1078;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f1077;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs147, %f1076;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs154, 0;
|
|
st.v4.u16 [%rd271], {%rs147, %rs148, %rs149, %rs154};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_160:
|
|
mov.u64 %rd284, image_RNM2;
|
|
cvta.global.u64 %rd279, %rd284;
|
|
mov.u32 %r388, 8;
|
|
mov.u64 %rd283, 0;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r107, %r388, %rd22, %rd23, %rd283, %rd283);
|
|
// inline asm
|
|
mov.f32 %f1079, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f1079;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs156, 0;
|
|
st.v4.u16 [%rd278], {%rs155, %rs155, %rs155, %rs156};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_87:
|
|
setp.geu.f32 %p86, %f1155, 0f00000000;
|
|
@%p86 bra BB0_90;
|
|
|
|
cvt.rzi.f32.f32 %f746, %f718;
|
|
setp.neu.f32 %p87, %f746, 0f3EE8BA2E;
|
|
selp.f32 %f1158, 0f7FFFFFFF, %f1158, %p87;
|
|
|
|
BB0_90:
|
|
add.f32 %f748, %f182, 0f3EE8BA2E;
|
|
mov.b32 %r289, %f748;
|
|
setp.lt.s32 %p89, %r289, 2139095040;
|
|
@%p89 bra BB0_95;
|
|
|
|
setp.gtu.f32 %p90, %f182, 0f7F800000;
|
|
@%p90 bra BB0_94;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_94:
|
|
add.f32 %f1158, %f1155, 0f3EE8BA2E;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_92:
|
|
setp.neu.f32 %p91, %f182, 0f7F800000;
|
|
@%p91 bra BB0_95;
|
|
|
|
selp.f32 %f1158, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_95:
|
|
mul.f32 %f749, %f1158, 0f437F0000;
|
|
setp.eq.f32 %p92, %f1155, 0f3F800000;
|
|
selp.f32 %f750, 0f437F0000, %f749, %p92;
|
|
cvt.rzi.u32.f32 %r290, %f750;
|
|
cvt.u16.u32 %rs40, %r290;
|
|
mov.u16 %rs41, 255;
|
|
st.v2.u8 [%rd97], {%rs40, %rs41};
|
|
ld.global.u32 %r432, [imageEnabled];
|
|
|
|
BB0_96:
|
|
ld.global.f32 %f751, [lightColor];
|
|
mul.f32 %f194, %f1151, %f751;
|
|
ld.global.f32 %f752, [lightColor+4];
|
|
mul.f32 %f195, %f1151, %f752;
|
|
ld.global.f32 %f753, [lightColor+8];
|
|
mul.f32 %f196, %f1151, %f753;
|
|
and.b32 %r291, %r432, 1;
|
|
setp.eq.b32 %p93, %r291, 1;
|
|
@!%p93 bra BB0_131;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_97:
|
|
mov.f32 %f756, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f757, %f756;
|
|
fma.rn.f32 %f758, %f757, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f197, %f758;
|
|
abs.f32 %f198, %f194;
|
|
setp.lt.f32 %p94, %f198, 0f00800000;
|
|
mul.f32 %f759, %f198, 0f4B800000;
|
|
selp.f32 %f760, 0fC3170000, 0fC2FE0000, %p94;
|
|
selp.f32 %f761, %f759, %f198, %p94;
|
|
mov.b32 %r292, %f761;
|
|
and.b32 %r293, %r292, 8388607;
|
|
or.b32 %r294, %r293, 1065353216;
|
|
mov.b32 %f762, %r294;
|
|
shr.u32 %r295, %r292, 23;
|
|
cvt.rn.f32.u32 %f763, %r295;
|
|
add.f32 %f764, %f760, %f763;
|
|
setp.gt.f32 %p95, %f762, 0f3FB504F3;
|
|
mul.f32 %f765, %f762, 0f3F000000;
|
|
add.f32 %f766, %f764, 0f3F800000;
|
|
selp.f32 %f767, %f765, %f762, %p95;
|
|
selp.f32 %f768, %f766, %f764, %p95;
|
|
add.f32 %f769, %f767, 0fBF800000;
|
|
add.f32 %f755, %f767, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f754,%f755;
|
|
// inline asm
|
|
add.f32 %f770, %f769, %f769;
|
|
mul.f32 %f771, %f754, %f770;
|
|
mul.f32 %f772, %f771, %f771;
|
|
mov.f32 %f773, 0f3C4CAF63;
|
|
mov.f32 %f774, 0f3B18F0FE;
|
|
fma.rn.f32 %f775, %f774, %f772, %f773;
|
|
mov.f32 %f776, 0f3DAAAABD;
|
|
fma.rn.f32 %f777, %f775, %f772, %f776;
|
|
mul.rn.f32 %f778, %f777, %f772;
|
|
mul.rn.f32 %f779, %f778, %f771;
|
|
sub.f32 %f780, %f769, %f771;
|
|
neg.f32 %f781, %f771;
|
|
add.f32 %f782, %f780, %f780;
|
|
fma.rn.f32 %f783, %f781, %f769, %f782;
|
|
mul.rn.f32 %f784, %f754, %f783;
|
|
add.f32 %f785, %f779, %f771;
|
|
sub.f32 %f786, %f771, %f785;
|
|
add.f32 %f787, %f779, %f786;
|
|
add.f32 %f788, %f784, %f787;
|
|
add.f32 %f789, %f785, %f788;
|
|
sub.f32 %f790, %f785, %f789;
|
|
add.f32 %f791, %f788, %f790;
|
|
mov.f32 %f792, 0f3F317200;
|
|
mul.rn.f32 %f793, %f768, %f792;
|
|
mov.f32 %f794, 0f35BFBE8E;
|
|
mul.rn.f32 %f795, %f768, %f794;
|
|
add.f32 %f796, %f793, %f789;
|
|
sub.f32 %f797, %f793, %f796;
|
|
add.f32 %f798, %f789, %f797;
|
|
add.f32 %f799, %f791, %f798;
|
|
add.f32 %f800, %f795, %f799;
|
|
add.f32 %f801, %f796, %f800;
|
|
sub.f32 %f802, %f796, %f801;
|
|
add.f32 %f803, %f800, %f802;
|
|
mov.f32 %f804, 0f3EE66666;
|
|
mul.rn.f32 %f805, %f804, %f801;
|
|
neg.f32 %f806, %f805;
|
|
fma.rn.f32 %f807, %f804, %f801, %f806;
|
|
fma.rn.f32 %f808, %f804, %f803, %f807;
|
|
mov.f32 %f809, 0f00000000;
|
|
fma.rn.f32 %f810, %f809, %f801, %f808;
|
|
add.rn.f32 %f811, %f805, %f810;
|
|
neg.f32 %f812, %f811;
|
|
add.rn.f32 %f813, %f805, %f812;
|
|
add.rn.f32 %f814, %f813, %f810;
|
|
mov.b32 %r296, %f811;
|
|
setp.eq.s32 %p96, %r296, 1118925336;
|
|
add.s32 %r297, %r296, -1;
|
|
mov.b32 %f815, %r297;
|
|
add.f32 %f816, %f814, 0f37000000;
|
|
selp.f32 %f817, %f815, %f811, %p96;
|
|
selp.f32 %f199, %f816, %f814, %p96;
|
|
mul.f32 %f818, %f817, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f819, %f818;
|
|
mov.f32 %f820, 0fBF317200;
|
|
fma.rn.f32 %f821, %f819, %f820, %f817;
|
|
mov.f32 %f822, 0fB5BFBE8E;
|
|
fma.rn.f32 %f823, %f819, %f822, %f821;
|
|
mul.f32 %f824, %f823, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f825, %f824;
|
|
add.f32 %f826, %f819, 0f00000000;
|
|
ex2.approx.f32 %f827, %f826;
|
|
mul.f32 %f828, %f825, %f827;
|
|
setp.lt.f32 %p97, %f817, 0fC2D20000;
|
|
selp.f32 %f829, 0f00000000, %f828, %p97;
|
|
setp.gt.f32 %p98, %f817, 0f42D20000;
|
|
selp.f32 %f1159, 0f7F800000, %f829, %p98;
|
|
setp.eq.f32 %p99, %f1159, 0f7F800000;
|
|
@%p99 bra BB0_99;
|
|
|
|
fma.rn.f32 %f1159, %f1159, %f199, %f1159;
|
|
|
|
BB0_99:
|
|
setp.lt.f32 %p100, %f194, 0f00000000;
|
|
setp.eq.f32 %p101, %f197, 0f3F800000;
|
|
and.pred %p4, %p100, %p101;
|
|
mov.b32 %r298, %f1159;
|
|
xor.b32 %r299, %r298, -2147483648;
|
|
mov.b32 %f830, %r299;
|
|
selp.f32 %f1161, %f830, %f1159, %p4;
|
|
setp.eq.f32 %p102, %f194, 0f00000000;
|
|
@%p102 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f833, %f194, %f194;
|
|
selp.f32 %f1161, %f833, 0f00000000, %p101;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.geu.f32 %p103, %f194, 0f00000000;
|
|
@%p103 bra BB0_103;
|
|
|
|
cvt.rzi.f32.f32 %f832, %f804;
|
|
setp.neu.f32 %p104, %f832, 0f3EE66666;
|
|
selp.f32 %f1161, 0f7FFFFFFF, %f1161, %p104;
|
|
|
|
BB0_103:
|
|
add.f32 %f834, %f198, 0f3EE66666;
|
|
mov.b32 %r300, %f834;
|
|
setp.lt.s32 %p106, %r300, 2139095040;
|
|
@%p106 bra BB0_108;
|
|
|
|
setp.gtu.f32 %p107, %f198, 0f7F800000;
|
|
@%p107 bra BB0_107;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_107:
|
|
add.f32 %f1161, %f194, 0f3EE66666;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_105:
|
|
setp.neu.f32 %p108, %f198, 0f7F800000;
|
|
@%p108 bra BB0_108;
|
|
|
|
selp.f32 %f1161, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_108:
|
|
setp.eq.f32 %p109, %f194, 0f3F800000;
|
|
selp.f32 %f210, 0f3F800000, %f1161, %p109;
|
|
abs.f32 %f211, %f195;
|
|
setp.lt.f32 %p110, %f211, 0f00800000;
|
|
mul.f32 %f837, %f211, 0f4B800000;
|
|
selp.f32 %f838, 0fC3170000, 0fC2FE0000, %p110;
|
|
selp.f32 %f839, %f837, %f211, %p110;
|
|
mov.b32 %r301, %f839;
|
|
and.b32 %r302, %r301, 8388607;
|
|
or.b32 %r303, %r302, 1065353216;
|
|
mov.b32 %f840, %r303;
|
|
shr.u32 %r304, %r301, 23;
|
|
cvt.rn.f32.u32 %f841, %r304;
|
|
add.f32 %f842, %f838, %f841;
|
|
setp.gt.f32 %p111, %f840, 0f3FB504F3;
|
|
mul.f32 %f843, %f840, 0f3F000000;
|
|
add.f32 %f844, %f842, 0f3F800000;
|
|
selp.f32 %f845, %f843, %f840, %p111;
|
|
selp.f32 %f846, %f844, %f842, %p111;
|
|
add.f32 %f847, %f845, 0fBF800000;
|
|
add.f32 %f836, %f845, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f835,%f836;
|
|
// inline asm
|
|
add.f32 %f848, %f847, %f847;
|
|
mul.f32 %f849, %f835, %f848;
|
|
mul.f32 %f850, %f849, %f849;
|
|
fma.rn.f32 %f853, %f774, %f850, %f773;
|
|
fma.rn.f32 %f855, %f853, %f850, %f776;
|
|
mul.rn.f32 %f856, %f855, %f850;
|
|
mul.rn.f32 %f857, %f856, %f849;
|
|
sub.f32 %f858, %f847, %f849;
|
|
neg.f32 %f859, %f849;
|
|
add.f32 %f860, %f858, %f858;
|
|
fma.rn.f32 %f861, %f859, %f847, %f860;
|
|
mul.rn.f32 %f862, %f835, %f861;
|
|
add.f32 %f863, %f857, %f849;
|
|
sub.f32 %f864, %f849, %f863;
|
|
add.f32 %f865, %f857, %f864;
|
|
add.f32 %f866, %f862, %f865;
|
|
add.f32 %f867, %f863, %f866;
|
|
sub.f32 %f868, %f863, %f867;
|
|
add.f32 %f869, %f866, %f868;
|
|
mul.rn.f32 %f871, %f846, %f792;
|
|
mul.rn.f32 %f873, %f846, %f794;
|
|
add.f32 %f874, %f871, %f867;
|
|
sub.f32 %f875, %f871, %f874;
|
|
add.f32 %f876, %f867, %f875;
|
|
add.f32 %f877, %f869, %f876;
|
|
add.f32 %f878, %f873, %f877;
|
|
add.f32 %f879, %f874, %f878;
|
|
sub.f32 %f880, %f874, %f879;
|
|
add.f32 %f881, %f878, %f880;
|
|
mul.rn.f32 %f883, %f804, %f879;
|
|
neg.f32 %f884, %f883;
|
|
fma.rn.f32 %f885, %f804, %f879, %f884;
|
|
fma.rn.f32 %f886, %f804, %f881, %f885;
|
|
fma.rn.f32 %f888, %f809, %f879, %f886;
|
|
add.rn.f32 %f889, %f883, %f888;
|
|
neg.f32 %f890, %f889;
|
|
add.rn.f32 %f891, %f883, %f890;
|
|
add.rn.f32 %f892, %f891, %f888;
|
|
mov.b32 %r305, %f889;
|
|
setp.eq.s32 %p112, %r305, 1118925336;
|
|
add.s32 %r306, %r305, -1;
|
|
mov.b32 %f893, %r306;
|
|
add.f32 %f894, %f892, 0f37000000;
|
|
selp.f32 %f895, %f893, %f889, %p112;
|
|
selp.f32 %f212, %f894, %f892, %p112;
|
|
mul.f32 %f896, %f895, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f897, %f896;
|
|
fma.rn.f32 %f899, %f897, %f820, %f895;
|
|
fma.rn.f32 %f901, %f897, %f822, %f899;
|
|
mul.f32 %f902, %f901, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f903, %f902;
|
|
add.f32 %f904, %f897, 0f00000000;
|
|
ex2.approx.f32 %f905, %f904;
|
|
mul.f32 %f906, %f903, %f905;
|
|
setp.lt.f32 %p113, %f895, 0fC2D20000;
|
|
selp.f32 %f907, 0f00000000, %f906, %p113;
|
|
setp.gt.f32 %p114, %f895, 0f42D20000;
|
|
selp.f32 %f1162, 0f7F800000, %f907, %p114;
|
|
setp.eq.f32 %p115, %f1162, 0f7F800000;
|
|
@%p115 bra BB0_110;
|
|
|
|
fma.rn.f32 %f1162, %f1162, %f212, %f1162;
|
|
|
|
BB0_110:
|
|
setp.lt.f32 %p116, %f195, 0f00000000;
|
|
and.pred %p5, %p116, %p101;
|
|
mov.b32 %r307, %f1162;
|
|
xor.b32 %r308, %r307, -2147483648;
|
|
mov.b32 %f908, %r308;
|
|
selp.f32 %f1164, %f908, %f1162, %p5;
|
|
setp.eq.f32 %p118, %f195, 0f00000000;
|
|
@%p118 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f911, %f195, %f195;
|
|
selp.f32 %f1164, %f911, 0f00000000, %p101;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.geu.f32 %p119, %f195, 0f00000000;
|
|
@%p119 bra BB0_114;
|
|
|
|
cvt.rzi.f32.f32 %f910, %f804;
|
|
setp.neu.f32 %p120, %f910, 0f3EE66666;
|
|
selp.f32 %f1164, 0f7FFFFFFF, %f1164, %p120;
|
|
|
|
BB0_114:
|
|
add.f32 %f912, %f211, 0f3EE66666;
|
|
mov.b32 %r309, %f912;
|
|
setp.lt.s32 %p122, %r309, 2139095040;
|
|
@%p122 bra BB0_119;
|
|
|
|
setp.gtu.f32 %p123, %f211, 0f7F800000;
|
|
@%p123 bra BB0_118;
|
|
bra.uni BB0_116;
|
|
|
|
BB0_118:
|
|
add.f32 %f1164, %f195, 0f3EE66666;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_116:
|
|
setp.neu.f32 %p124, %f211, 0f7F800000;
|
|
@%p124 bra BB0_119;
|
|
|
|
selp.f32 %f1164, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_119:
|
|
setp.eq.f32 %p125, %f195, 0f3F800000;
|
|
selp.f32 %f223, 0f3F800000, %f1164, %p125;
|
|
abs.f32 %f224, %f196;
|
|
setp.lt.f32 %p126, %f224, 0f00800000;
|
|
mul.f32 %f915, %f224, 0f4B800000;
|
|
selp.f32 %f916, 0fC3170000, 0fC2FE0000, %p126;
|
|
selp.f32 %f917, %f915, %f224, %p126;
|
|
mov.b32 %r310, %f917;
|
|
and.b32 %r311, %r310, 8388607;
|
|
or.b32 %r312, %r311, 1065353216;
|
|
mov.b32 %f918, %r312;
|
|
shr.u32 %r313, %r310, 23;
|
|
cvt.rn.f32.u32 %f919, %r313;
|
|
add.f32 %f920, %f916, %f919;
|
|
setp.gt.f32 %p127, %f918, 0f3FB504F3;
|
|
mul.f32 %f921, %f918, 0f3F000000;
|
|
add.f32 %f922, %f920, 0f3F800000;
|
|
selp.f32 %f923, %f921, %f918, %p127;
|
|
selp.f32 %f924, %f922, %f920, %p127;
|
|
add.f32 %f925, %f923, 0fBF800000;
|
|
add.f32 %f914, %f923, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f913,%f914;
|
|
// inline asm
|
|
add.f32 %f926, %f925, %f925;
|
|
mul.f32 %f927, %f913, %f926;
|
|
mul.f32 %f928, %f927, %f927;
|
|
fma.rn.f32 %f931, %f774, %f928, %f773;
|
|
fma.rn.f32 %f933, %f931, %f928, %f776;
|
|
mul.rn.f32 %f934, %f933, %f928;
|
|
mul.rn.f32 %f935, %f934, %f927;
|
|
sub.f32 %f936, %f925, %f927;
|
|
neg.f32 %f937, %f927;
|
|
add.f32 %f938, %f936, %f936;
|
|
fma.rn.f32 %f939, %f937, %f925, %f938;
|
|
mul.rn.f32 %f940, %f913, %f939;
|
|
add.f32 %f941, %f935, %f927;
|
|
sub.f32 %f942, %f927, %f941;
|
|
add.f32 %f943, %f935, %f942;
|
|
add.f32 %f944, %f940, %f943;
|
|
add.f32 %f945, %f941, %f944;
|
|
sub.f32 %f946, %f941, %f945;
|
|
add.f32 %f947, %f944, %f946;
|
|
mul.rn.f32 %f949, %f924, %f792;
|
|
mul.rn.f32 %f951, %f924, %f794;
|
|
add.f32 %f952, %f949, %f945;
|
|
sub.f32 %f953, %f949, %f952;
|
|
add.f32 %f954, %f945, %f953;
|
|
add.f32 %f955, %f947, %f954;
|
|
add.f32 %f956, %f951, %f955;
|
|
add.f32 %f957, %f952, %f956;
|
|
sub.f32 %f958, %f952, %f957;
|
|
add.f32 %f959, %f956, %f958;
|
|
mul.rn.f32 %f961, %f804, %f957;
|
|
neg.f32 %f962, %f961;
|
|
fma.rn.f32 %f963, %f804, %f957, %f962;
|
|
fma.rn.f32 %f964, %f804, %f959, %f963;
|
|
fma.rn.f32 %f966, %f809, %f957, %f964;
|
|
add.rn.f32 %f967, %f961, %f966;
|
|
neg.f32 %f968, %f967;
|
|
add.rn.f32 %f969, %f961, %f968;
|
|
add.rn.f32 %f970, %f969, %f966;
|
|
mov.b32 %r314, %f967;
|
|
setp.eq.s32 %p128, %r314, 1118925336;
|
|
add.s32 %r315, %r314, -1;
|
|
mov.b32 %f971, %r315;
|
|
add.f32 %f972, %f970, 0f37000000;
|
|
selp.f32 %f973, %f971, %f967, %p128;
|
|
selp.f32 %f225, %f972, %f970, %p128;
|
|
mul.f32 %f974, %f973, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f975, %f974;
|
|
fma.rn.f32 %f977, %f975, %f820, %f973;
|
|
fma.rn.f32 %f979, %f975, %f822, %f977;
|
|
mul.f32 %f980, %f979, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f981, %f980;
|
|
add.f32 %f982, %f975, 0f00000000;
|
|
ex2.approx.f32 %f983, %f982;
|
|
mul.f32 %f984, %f981, %f983;
|
|
setp.lt.f32 %p129, %f973, 0fC2D20000;
|
|
selp.f32 %f985, 0f00000000, %f984, %p129;
|
|
setp.gt.f32 %p130, %f973, 0f42D20000;
|
|
selp.f32 %f1165, 0f7F800000, %f985, %p130;
|
|
setp.eq.f32 %p131, %f1165, 0f7F800000;
|
|
@%p131 bra BB0_121;
|
|
|
|
fma.rn.f32 %f1165, %f1165, %f225, %f1165;
|
|
|
|
BB0_121:
|
|
setp.lt.f32 %p132, %f196, 0f00000000;
|
|
and.pred %p6, %p132, %p101;
|
|
mov.b32 %r316, %f1165;
|
|
xor.b32 %r317, %r316, -2147483648;
|
|
mov.b32 %f986, %r317;
|
|
selp.f32 %f1167, %f986, %f1165, %p6;
|
|
setp.eq.f32 %p134, %f196, 0f00000000;
|
|
@%p134 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f989, %f196, %f196;
|
|
selp.f32 %f1167, %f989, 0f00000000, %p101;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.geu.f32 %p135, %f196, 0f00000000;
|
|
@%p135 bra BB0_125;
|
|
|
|
cvt.rzi.f32.f32 %f988, %f804;
|
|
setp.neu.f32 %p136, %f988, 0f3EE66666;
|
|
selp.f32 %f1167, 0f7FFFFFFF, %f1167, %p136;
|
|
|
|
BB0_125:
|
|
add.f32 %f990, %f224, 0f3EE66666;
|
|
mov.b32 %r318, %f990;
|
|
setp.lt.s32 %p138, %r318, 2139095040;
|
|
@%p138 bra BB0_130;
|
|
|
|
setp.gtu.f32 %p139, %f224, 0f7F800000;
|
|
@%p139 bra BB0_129;
|
|
bra.uni BB0_127;
|
|
|
|
BB0_129:
|
|
add.f32 %f1167, %f196, 0f3EE66666;
|
|
bra.uni BB0_130;
|
|
|
|
BB0_127:
|
|
setp.neu.f32 %p140, %f224, 0f7F800000;
|
|
@%p140 bra BB0_130;
|
|
|
|
selp.f32 %f1167, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_130:
|
|
mov.u32 %r392, 2;
|
|
setp.eq.f32 %p141, %f196, 0f3F800000;
|
|
selp.f32 %f991, 0f3F800000, %f1167, %p141;
|
|
cvt.u64.u32 %rd107, %r4;
|
|
cvt.u64.u32 %rd106, %r3;
|
|
mov.u64 %rd110, image;
|
|
cvta.global.u64 %rd105, %rd110;
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd105, %r392, %r108, %rd106, %rd107, %rd29, %rd29);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f992, %f991;
|
|
mul.f32 %f993, %f992, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r321, %f993;
|
|
cvt.sat.f32.f32 %f994, %f223;
|
|
mul.f32 %f995, %f994, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r322, %f995;
|
|
cvt.sat.f32.f32 %f996, %f210;
|
|
mul.f32 %f997, %f996, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r323, %f997;
|
|
cvt.u16.u32 %rs42, %r321;
|
|
cvt.u16.u32 %rs43, %r323;
|
|
cvt.u16.u32 %rs44, %r322;
|
|
mov.u16 %rs45, 255;
|
|
st.v4.u8 [%rd104], {%rs42, %rs44, %rs43, %rs45};
|
|
ld.global.u32 %r432, [imageEnabled];
|
|
|
|
BB0_131:
|
|
cvt.u64.u32 %rd20, %r3;
|
|
cvt.u64.u32 %rd21, %r4;
|
|
and.b32 %r324, %r432, 4;
|
|
setp.eq.s32 %p142, %r324, 0;
|
|
@%p142 bra BB0_135;
|
|
|
|
ld.global.u32 %r325, [additive];
|
|
setp.eq.s32 %p143, %r325, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f539;}
|
|
|
|
// inline asm
|
|
@%p143 bra BB0_134;
|
|
|
|
mov.u32 %r393, 2;
|
|
mov.u64 %rd123, image_HDR;
|
|
cvta.global.u64 %rd112, %rd123;
|
|
mov.u32 %r329, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r393, %r329, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
ld.v4.u16 {%rs53, %rs54, %rs55, %rs56}, [%rd111];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f999, %rs53;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1000, %rs54;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1001, %rs55;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd112, %r393, %r329, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
add.f32 %f1002, %f194, %f999;
|
|
add.f32 %f1003, %f195, %f1000;
|
|
add.f32 %f1004, %f196, %f1001;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f1004;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f1003;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f1002;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd117], {%rs50, %rs51, %rs52, %rs46};
|
|
bra.uni BB0_135;
|
|
|
|
BB0_134:
|
|
mov.u32 %r394, 2;
|
|
mov.u64 %rd130, image_HDR;
|
|
cvta.global.u64 %rd125, %rd130;
|
|
mov.u32 %r331, 8;
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd125, %r394, %r331, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f196;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f195;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f194;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs57, %rs58, %rs59, %rs46};
|
|
|
|
BB0_135:
|
|
add.f32 %f1009, %f1152, %f1153;
|
|
add.f32 %f1010, %f1009, %f1154;
|
|
mul.f32 %f1011, %f1010, 0f3F13CD3A;
|
|
div.rn.f32 %f1012, %f1151, %f1011;
|
|
setp.eq.f32 %p144, %f1151, 0f00000000;
|
|
selp.f32 %f1013, 0f00000000, %f1012, %p144;
|
|
mul.f32 %f1014, %f1152, %f1013;
|
|
mul.f32 %f1015, %f1153, %f1013;
|
|
mul.f32 %f1016, %f1154, %f1013;
|
|
ld.global.f32 %f1017, [lightColor];
|
|
mul.f32 %f236, %f1017, %f1014;
|
|
ld.global.f32 %f1018, [lightColor+4];
|
|
mul.f32 %f237, %f1018, %f1014;
|
|
ld.global.f32 %f1019, [lightColor+8];
|
|
mul.f32 %f238, %f1014, %f1019;
|
|
mul.f32 %f239, %f1017, %f1015;
|
|
mul.f32 %f240, %f1018, %f1015;
|
|
mul.f32 %f241, %f1015, %f1019;
|
|
mul.f32 %f242, %f1016, %f1017;
|
|
mul.f32 %f243, %f1016, %f1018;
|
|
mul.f32 %f244, %f1016, %f1019;
|
|
ld.global.u32 %r332, [additive];
|
|
setp.eq.s32 %p145, %r332, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f539;}
|
|
|
|
// inline asm
|
|
@%p145 bra BB0_137;
|
|
|
|
mov.u32 %r395, 2;
|
|
mov.u64 %rd143, image_RNM0;
|
|
cvta.global.u64 %rd132, %rd143;
|
|
mov.u32 %r336, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r395, %r336, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd131];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1020, %rs67;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1021, %rs68;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1022, %rs69;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd132, %r395, %r336, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
add.f32 %f1023, %f236, %f1020;
|
|
add.f32 %f1024, %f237, %f1021;
|
|
add.f32 %f1025, %f238, %f1022;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f1025;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f1024;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f1023;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd137], {%rs64, %rs65, %rs66, %rs60};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_137:
|
|
mov.u32 %r400, 2;
|
|
mov.u64 %rd150, image_RNM0;
|
|
cvta.global.u64 %rd145, %rd150;
|
|
mov.u32 %r338, 8;
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd145, %r400, %r338, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f238;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f237;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f236;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs71, %rs72, %rs73, %rs60};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r339, [additive];
|
|
setp.eq.s32 %p146, %r339, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f539;}
|
|
|
|
// inline asm
|
|
@%p146 bra BB0_140;
|
|
|
|
mov.u32 %r396, 2;
|
|
mov.u64 %rd163, image_RNM1;
|
|
cvta.global.u64 %rd152, %rd163;
|
|
mov.u32 %r343, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r396, %r343, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
ld.v4.u16 {%rs81, %rs82, %rs83, %rs84}, [%rd151];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1030, %rs81;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1031, %rs82;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1032, %rs83;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd152, %r396, %r343, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
add.f32 %f1033, %f239, %f1030;
|
|
add.f32 %f1034, %f240, %f1031;
|
|
add.f32 %f1035, %f241, %f1032;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f1035;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f1034;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f1033;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd157], {%rs78, %rs79, %rs80, %rs74};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_140:
|
|
mov.u32 %r399, 2;
|
|
mov.u64 %rd170, image_RNM1;
|
|
cvta.global.u64 %rd165, %rd170;
|
|
mov.u32 %r345, 8;
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd165, %r399, %r345, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f239;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs85, %rs86, %rs87, %rs74};
|
|
|
|
BB0_141:
|
|
ld.global.u32 %r346, [additive];
|
|
setp.eq.s32 %p147, %r346, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f539;}
|
|
|
|
// inline asm
|
|
@%p147 bra BB0_143;
|
|
|
|
mov.u32 %r397, 2;
|
|
mov.u64 %rd183, image_RNM2;
|
|
cvta.global.u64 %rd172, %rd183;
|
|
mov.u32 %r350, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r397, %r350, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
ld.v4.u16 {%rs95, %rs96, %rs97, %rs98}, [%rd171];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1040, %rs95;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1041, %rs96;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1042, %rs97;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd172, %r397, %r350, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
add.f32 %f1043, %f242, %f1040;
|
|
add.f32 %f1044, %f243, %f1041;
|
|
add.f32 %f1045, %f244, %f1042;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f1045;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f1044;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f1043;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd177], {%rs92, %rs93, %rs94, %rs88};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_143:
|
|
mov.u32 %r398, 2;
|
|
mov.u64 %rd190, image_RNM2;
|
|
cvta.global.u64 %rd185, %rd190;
|
|
mov.u32 %r352, 8;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r398, %r352, %rd20, %rd21, %rd29, %rd29);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f244;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f243;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f242;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs99, %rs100, %rs101, %rs88};
|
|
|
|
BB0_161:
|
|
ret;
|
|
}
|
|
|
|
|