3170 lines
85 KiB
Plaintext
3170 lines
85 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 image_RNM3[1];
|
|
.global .align 1 .b8 image_RNM4[1];
|
|
.global .align 1 .b8 image_RNM5[1];
|
|
.global .align 1 .b8 image_RNM6[1];
|
|
.global .align 1 .b8 image_RNM7[1];
|
|
.global .align 1 .b8 image_RNM8[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[36];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<159>;
|
|
.reg .b16 %rs<299>;
|
|
.reg .f32 %f<1275>;
|
|
.reg .b32 %r<483>;
|
|
.reg .b64 %rd<519>;
|
|
|
|
|
|
mov.u64 %rd518, __local_depot0;
|
|
cvta.local.u64 %SP, %rd518;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
shl.b32 %r4, %r3, 1;
|
|
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
|
cvt.u64.u32 %rd24, %r103;
|
|
cvt.u64.u32 %rd25, %r104;
|
|
mov.u64 %rd28, uvnormal;
|
|
cvta.global.u64 %rd23, %rd28;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd27, 0;
|
|
// inline asm
|
|
call (%rd22), _rt_buffer_get_64, (%rd23, %r101, %r102, %rd24, %rd25, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd22];
|
|
shr.u32 %r107, %r5, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r5;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p7, %rs14, 0;
|
|
mov.f32 %f1168, 0f00000000;
|
|
mov.f32 %f1169, %f1168;
|
|
mov.f32 %f1170, %f1168;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd22+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f298, %rs17;
|
|
div.rn.f32 %f299, %f298, 0f437F0000;
|
|
fma.rn.f32 %f300, %f299, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f301, %rs15;
|
|
div.rn.f32 %f302, %f301, 0f437F0000;
|
|
fma.rn.f32 %f303, %f302, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f304, %rs12;
|
|
div.rn.f32 %f305, %f304, 0f437F0000;
|
|
fma.rn.f32 %f306, %f305, 0f40000000, 0fBF800000;
|
|
mul.f32 %f307, %f303, %f303;
|
|
fma.rn.f32 %f308, %f300, %f300, %f307;
|
|
fma.rn.f32 %f309, %f306, %f306, %f308;
|
|
sqrt.rn.f32 %f310, %f309;
|
|
rcp.rn.f32 %f311, %f310;
|
|
mul.f32 %f1168, %f300, %f311;
|
|
mul.f32 %f1169, %f303, %f311;
|
|
mul.f32 %f1170, %f306, %f311;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
|
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
|
|
add.s32 %r6, %r108, %r111;
|
|
add.s32 %r7, %r109, %r112;
|
|
setp.eq.f32 %p8, %f1169, 0f00000000;
|
|
setp.eq.f32 %p9, %f1168, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1170, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_155;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_155:
|
|
ld.global.u32 %r482, [imageEnabled];
|
|
and.b32 %r372, %r482, 1;
|
|
setp.eq.b32 %p145, %r372, 1;
|
|
@!%p145 bra BB0_157;
|
|
bra.uni BB0_156;
|
|
|
|
BB0_156:
|
|
cvt.u64.u32 %rd299, %r6;
|
|
cvt.u64.u32 %rd300, %r7;
|
|
mov.u64 %rd303, image;
|
|
cvta.global.u64 %rd298, %rd303;
|
|
mov.u64 %rd302, 0;
|
|
// inline asm
|
|
call (%rd297), _rt_buffer_get_64, (%rd298, %r101, %r102, %rd299, %rd300, %rd302, %rd302);
|
|
// inline asm
|
|
mov.u16 %rs166, 0;
|
|
st.v4.u8 [%rd297], {%rs166, %rs166, %rs166, %rs166};
|
|
ld.global.u32 %r482, [imageEnabled];
|
|
|
|
BB0_157:
|
|
and.b32 %r375, %r482, 8;
|
|
setp.eq.s32 %p146, %r375, 0;
|
|
@%p146 bra BB0_159;
|
|
|
|
cvt.u64.u32 %rd307, %r7;
|
|
cvt.u64.u32 %rd306, %r6;
|
|
mov.u64 %rd310, image_Mask;
|
|
cvta.global.u64 %rd305, %rd310;
|
|
mov.u64 %rd309, 0;
|
|
// inline asm
|
|
call (%rd304), _rt_buffer_get_64, (%rd305, %r101, %r101, %rd306, %rd307, %rd309, %rd309);
|
|
// inline asm
|
|
mov.f32 %f1090, 0f00000000;
|
|
cvt.rzi.u32.f32 %r378, %f1090;
|
|
cvt.u16.u32 %rs167, %r378;
|
|
mov.u16 %rs168, 0;
|
|
st.v2.u8 [%rd304], {%rs167, %rs168};
|
|
ld.global.u32 %r482, [imageEnabled];
|
|
|
|
BB0_159:
|
|
cvt.u64.u32 %rd20, %r6;
|
|
cvt.u64.u32 %rd21, %r7;
|
|
and.b32 %r379, %r482, 4;
|
|
setp.eq.s32 %p147, %r379, 0;
|
|
@%p147 bra BB0_163;
|
|
|
|
ld.global.u32 %r380, [additive];
|
|
setp.eq.s32 %p148, %r380, 0;
|
|
@%p148 bra BB0_162;
|
|
|
|
mov.u64 %rd323, image_HDR;
|
|
cvta.global.u64 %rd312, %rd323;
|
|
mov.u32 %r384, 8;
|
|
mov.u64 %rd322, 0;
|
|
// inline asm
|
|
call (%rd311), _rt_buffer_get_64, (%rd312, %r101, %r384, %rd20, %rd21, %rd322, %rd322);
|
|
// inline asm
|
|
ld.v4.u16 {%rs175, %rs176, %rs177, %rs178}, [%rd311];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1091, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1092, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1093, %rs177;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd317), _rt_buffer_get_64, (%rd312, %r101, %r384, %rd20, %rd21, %rd322, %rd322);
|
|
// inline asm
|
|
add.f32 %f1094, %f1091, 0f00000000;
|
|
add.f32 %f1095, %f1092, 0f00000000;
|
|
add.f32 %f1096, %f1093, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f1096;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f1095;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f1094;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs179, 0;
|
|
st.v4.u16 [%rd317], {%rs172, %rs173, %rs174, %rs179};
|
|
bra.uni BB0_163;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd31, %r121;
|
|
cvt.u64.u32 %rd32, %r122;
|
|
mov.u64 %rd41, uvpos;
|
|
cvta.global.u64 %rd30, %rd41;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r101, %r117, %rd31, %rd32, %rd27, %rd27);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd29+8];
|
|
ld.f32 %f8, [%rd29+4];
|
|
ld.f32 %f7, [%rd29];
|
|
mul.f32 %f323, %f7, 0f3456BF95;
|
|
mul.f32 %f324, %f8, 0f3456BF95;
|
|
mul.f32 %f325, %f9, 0f3456BF95;
|
|
abs.f32 %f326, %f1168;
|
|
div.rn.f32 %f327, %f323, %f326;
|
|
abs.f32 %f328, %f1169;
|
|
div.rn.f32 %f329, %f324, %f328;
|
|
abs.f32 %f330, %f1170;
|
|
div.rn.f32 %f331, %f325, %f330;
|
|
abs.f32 %f332, %f327;
|
|
abs.f32 %f333, %f329;
|
|
abs.f32 %f334, %f331;
|
|
mov.f32 %f335, 0f38D1B717;
|
|
max.f32 %f336, %f332, %f335;
|
|
max.f32 %f337, %f333, %f335;
|
|
max.f32 %f338, %f334, %f335;
|
|
fma.rn.f32 %f10, %f1168, %f336, %f7;
|
|
fma.rn.f32 %f11, %f1169, %f337, %f8;
|
|
fma.rn.f32 %f12, %f1170, %f338, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd37, %r125;
|
|
cvt.u64.u32 %rd38, %r126;
|
|
mov.u64 %rd42, rnd_seeds;
|
|
cvta.global.u64 %rd36, %rd42;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd36, %r101, %r102, %rd37, %rd38, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r458, [%rd35];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r454, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f13, [lightPointSize];
|
|
mul.f32 %f14, %f10, 0f3456BF95;
|
|
mul.f32 %f15, %f11, 0f3456BF95;
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r454, 0;
|
|
abs.f32 %f472, %f15;
|
|
abs.f32 %f473, %f14;
|
|
max.f32 %f474, %f473, %f472;
|
|
abs.f32 %f475, %f16;
|
|
max.f32 %f476, %f474, %f475;
|
|
mov.f32 %f18, %f17;
|
|
mov.f32 %f19, %f17;
|
|
mov.f32 %f20, %f17;
|
|
mov.f32 %f21, %f17;
|
|
mov.f32 %f22, %f17;
|
|
mov.f32 %f23, %f17;
|
|
mov.f32 %f24, %f17;
|
|
mov.f32 %f25, %f17;
|
|
mov.f32 %f26, %f17;
|
|
mov.f32 %f27, %f17;
|
|
|
|
BB0_6:
|
|
shl.b32 %r10, %r454, 1;
|
|
cvt.s64.s32 %rd45, %r10;
|
|
mov.u64 %rd55, lightMeshBuffer;
|
|
cvta.global.u64 %rd44, %rd55;
|
|
mov.u32 %r130, 1;
|
|
mov.u64 %rd54, 0;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r130, %r117, %rd45, %rd54, %rd54, %rd54);
|
|
// inline asm
|
|
ld.f32 %f350, [%rd43];
|
|
sub.f32 %f351, %f350, %f7;
|
|
ld.f32 %f352, [%rd43+4];
|
|
sub.f32 %f353, %f352, %f8;
|
|
ld.f32 %f354, [%rd43+8];
|
|
sub.f32 %f355, %f354, %f9;
|
|
mul.f32 %f356, %f353, %f353;
|
|
fma.rn.f32 %f357, %f351, %f351, %f356;
|
|
fma.rn.f32 %f358, %f355, %f355, %f357;
|
|
sqrt.rn.f32 %f28, %f358;
|
|
rcp.rn.f32 %f359, %f28;
|
|
mul.f32 %f29, %f351, %f359;
|
|
mul.f32 %f30, %f353, %f359;
|
|
mul.f32 %f31, %f355, %f359;
|
|
mul.f32 %f360, %f28, %f28;
|
|
div.rn.f32 %f32, %f13, %f360;
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
|
cvt.u64.u32 %rd56, %r135;
|
|
cvt.u64.u32 %rd57, %r134;
|
|
and.b64 %rd51, %rd57, 255;
|
|
and.b64 %rd52, %rd56, 255;
|
|
mov.u64 %rd58, rnd_seeds2;
|
|
cvta.global.u64 %rd50, %rd58;
|
|
mov.u32 %r133, 3;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd50, %r101, %r133, %rd51, %rd52, %rd54, %rd54);
|
|
// inline asm
|
|
setp.eq.s32 %p15, %r3, 0;
|
|
@%p15 bra BB0_8;
|
|
|
|
ld.u8 %rs19, [%rd49];
|
|
cvt.rn.f32.u16 %f361, %rs19;
|
|
div.rn.f32 %f362, %f361, 0fC37F0000;
|
|
fma.rn.f32 %f363, %f362, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p17, %f32, %f363;
|
|
mov.pred %p158, -1;
|
|
@%p17 bra BB0_23;
|
|
|
|
BB0_8:
|
|
ld.global.f32 %f366, [lightInvCutoff];
|
|
mul.f32 %f33, %f28, %f366;
|
|
mov.f32 %f370, 0f40800000;
|
|
abs.f32 %f35, %f33;
|
|
setp.lt.f32 %p18, %f35, 0f00800000;
|
|
mul.f32 %f372, %f35, 0f4B800000;
|
|
selp.f32 %f373, 0fC3170000, 0fC2FE0000, %p18;
|
|
selp.f32 %f374, %f372, %f35, %p18;
|
|
mov.b32 %r138, %f374;
|
|
and.b32 %r139, %r138, 8388607;
|
|
or.b32 %r140, %r139, 1065353216;
|
|
mov.b32 %f375, %r140;
|
|
shr.u32 %r141, %r138, 23;
|
|
cvt.rn.f32.u32 %f376, %r141;
|
|
add.f32 %f377, %f373, %f376;
|
|
setp.gt.f32 %p19, %f375, 0f3FB504F3;
|
|
mul.f32 %f378, %f375, 0f3F000000;
|
|
add.f32 %f379, %f377, 0f3F800000;
|
|
selp.f32 %f380, %f378, %f375, %p19;
|
|
selp.f32 %f381, %f379, %f377, %p19;
|
|
add.f32 %f382, %f380, 0fBF800000;
|
|
add.f32 %f365, %f380, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f364,%f365;
|
|
// inline asm
|
|
add.f32 %f383, %f382, %f382;
|
|
mul.f32 %f384, %f364, %f383;
|
|
mul.f32 %f385, %f384, %f384;
|
|
mov.f32 %f386, 0f3C4CAF63;
|
|
mov.f32 %f387, 0f3B18F0FE;
|
|
fma.rn.f32 %f388, %f387, %f385, %f386;
|
|
mov.f32 %f389, 0f3DAAAABD;
|
|
fma.rn.f32 %f390, %f388, %f385, %f389;
|
|
mul.rn.f32 %f391, %f390, %f385;
|
|
mul.rn.f32 %f392, %f391, %f384;
|
|
sub.f32 %f393, %f382, %f384;
|
|
neg.f32 %f394, %f384;
|
|
add.f32 %f395, %f393, %f393;
|
|
fma.rn.f32 %f396, %f394, %f382, %f395;
|
|
mul.rn.f32 %f397, %f364, %f396;
|
|
add.f32 %f398, %f392, %f384;
|
|
sub.f32 %f399, %f384, %f398;
|
|
add.f32 %f400, %f392, %f399;
|
|
add.f32 %f401, %f397, %f400;
|
|
add.f32 %f402, %f398, %f401;
|
|
sub.f32 %f403, %f398, %f402;
|
|
add.f32 %f404, %f401, %f403;
|
|
mov.f32 %f405, 0f3F317200;
|
|
mul.rn.f32 %f406, %f381, %f405;
|
|
mov.f32 %f407, 0f35BFBE8E;
|
|
mul.rn.f32 %f408, %f381, %f407;
|
|
add.f32 %f409, %f406, %f402;
|
|
sub.f32 %f410, %f406, %f409;
|
|
add.f32 %f411, %f402, %f410;
|
|
add.f32 %f412, %f404, %f411;
|
|
add.f32 %f413, %f408, %f412;
|
|
add.f32 %f414, %f409, %f413;
|
|
sub.f32 %f415, %f409, %f414;
|
|
add.f32 %f416, %f413, %f415;
|
|
mul.rn.f32 %f417, %f370, %f414;
|
|
neg.f32 %f418, %f417;
|
|
fma.rn.f32 %f419, %f370, %f414, %f418;
|
|
fma.rn.f32 %f420, %f370, %f416, %f419;
|
|
mov.f32 %f421, 0f00000000;
|
|
fma.rn.f32 %f422, %f421, %f414, %f420;
|
|
add.rn.f32 %f423, %f417, %f422;
|
|
neg.f32 %f424, %f423;
|
|
add.rn.f32 %f425, %f417, %f424;
|
|
add.rn.f32 %f426, %f425, %f422;
|
|
mov.b32 %r142, %f423;
|
|
setp.eq.s32 %p20, %r142, 1118925336;
|
|
add.s32 %r143, %r142, -1;
|
|
mov.b32 %f427, %r143;
|
|
add.f32 %f428, %f426, 0f37000000;
|
|
selp.f32 %f429, %f427, %f423, %p20;
|
|
selp.f32 %f36, %f428, %f426, %p20;
|
|
mul.f32 %f430, %f429, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f431, %f430;
|
|
mov.f32 %f432, 0fBF317200;
|
|
fma.rn.f32 %f433, %f431, %f432, %f429;
|
|
mov.f32 %f434, 0fB5BFBE8E;
|
|
fma.rn.f32 %f435, %f431, %f434, %f433;
|
|
mul.f32 %f436, %f435, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f437, %f436;
|
|
add.f32 %f438, %f431, 0f00000000;
|
|
ex2.approx.f32 %f439, %f438;
|
|
mul.f32 %f440, %f437, %f439;
|
|
setp.lt.f32 %p21, %f429, 0fC2D20000;
|
|
selp.f32 %f441, 0f00000000, %f440, %p21;
|
|
setp.gt.f32 %p22, %f429, 0f42D20000;
|
|
selp.f32 %f1182, 0f7F800000, %f441, %p22;
|
|
setp.eq.f32 %p23, %f1182, 0f7F800000;
|
|
@%p23 bra BB0_10;
|
|
|
|
fma.rn.f32 %f1182, %f1182, %f36, %f1182;
|
|
|
|
BB0_10:
|
|
mov.f32 %f1166, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1165, %f1166;
|
|
add.f32 %f1164, %f1165, %f1165;
|
|
mov.f32 %f1163, 0f40800000;
|
|
sub.f32 %f1162, %f1163, %f1164;
|
|
abs.f32 %f1161, %f1162;
|
|
setp.lt.f32 %p24, %f33, 0f00000000;
|
|
setp.eq.f32 %p25, %f1161, 0f3F800000;
|
|
and.pred %p1, %p24, %p25;
|
|
mov.b32 %r144, %f1182;
|
|
xor.b32 %r145, %r144, -2147483648;
|
|
mov.b32 %f442, %r145;
|
|
selp.f32 %f1184, %f442, %f1182, %p1;
|
|
setp.eq.f32 %p26, %f33, 0f00000000;
|
|
@%p26 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f445, %f33, %f33;
|
|
selp.f32 %f1184, %f445, 0f00000000, %p25;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p27, %f33, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
|
|
mov.f32 %f1167, 0f40800000;
|
|
cvt.rzi.f32.f32 %f444, %f1167;
|
|
setp.neu.f32 %p28, %f444, 0f40800000;
|
|
selp.f32 %f1184, 0f7FFFFFFF, %f1184, %p28;
|
|
|
|
BB0_14:
|
|
add.f32 %f446, %f35, 0f40800000;
|
|
mov.b32 %r146, %f446;
|
|
setp.lt.s32 %p30, %r146, 2139095040;
|
|
@%p30 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p31, %f35, 0f7F800000;
|
|
@%p31 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f1184, %f33, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p32, %f35, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
|
|
selp.f32 %f1184, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u64 %rd513, 0;
|
|
mov.u32 %r451, 1;
|
|
mov.u64 %rd512, lightMeshBuffer;
|
|
cvta.global.u64 %rd511, %rd512;
|
|
shl.b32 %r450, %r454, 1;
|
|
mov.f32 %f447, 0f3F800000;
|
|
sub.f32 %f448, %f447, %f1184;
|
|
setp.eq.f32 %p33, %f33, 0f3F800000;
|
|
selp.f32 %f449, 0f00000000, %f448, %p33;
|
|
cvt.sat.f32.f32 %f450, %f449;
|
|
mul.f32 %f451, %f32, %f450;
|
|
add.s32 %r149, %r450, 1;
|
|
cvt.s64.s32 %rd61, %r149;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd511, %r451, %r117, %rd61, %rd513, %rd513, %rd513);
|
|
// inline asm
|
|
ld.f32 %f452, [%rd59];
|
|
mul.f32 %f453, %f29, %f452;
|
|
ld.f32 %f454, [%rd59+4];
|
|
mul.f32 %f455, %f30, %f454;
|
|
neg.f32 %f456, %f455;
|
|
sub.f32 %f457, %f456, %f453;
|
|
ld.f32 %f458, [%rd59+8];
|
|
mul.f32 %f459, %f31, %f458;
|
|
sub.f32 %f460, %f457, %f459;
|
|
cvt.sat.f32.f32 %f461, %f460;
|
|
mul.f32 %f47, %f451, %f461;
|
|
mul.f32 %f462, %f1169, %f30;
|
|
fma.rn.f32 %f463, %f1168, %f29, %f462;
|
|
fma.rn.f32 %f48, %f1170, %f31, %f463;
|
|
setp.leu.f32 %p34, %f47, 0f3727C5AC;
|
|
@%p34 bra BB0_21;
|
|
|
|
mov.u32 %r452, 1;
|
|
add.u64 %rd66, %SP, 4;
|
|
cvta.to.local.u64 %rd67, %rd66;
|
|
max.f32 %f470, %f476, %f335;
|
|
sub.f32 %f471, %f28, %f470;
|
|
mov.u32 %r153, 1065353216;
|
|
st.local.u32 [%rd67], %r153;
|
|
ld.global.u32 %r150, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r150, %f10, %f11, %f12, %f29, %f30, %f31, %r452, %f470, %f471, %rd66, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f478, [%rd67];
|
|
mul.f32 %f479, %f47, 0f3EA2F983;
|
|
mul.f32 %f480, %f479, %f478;
|
|
cvt.sat.f32.f32 %f481, %f48;
|
|
fma.rn.f32 %f27, %f481, %f480, %f27;
|
|
add.f32 %f26, %f26, %f480;
|
|
fma.rn.f32 %f25, %f30, %f480, %f25;
|
|
fma.rn.f32 %f24, %f31, %f480, %f24;
|
|
fma.rn.f32 %f23, %f29, %f480, %f23;
|
|
mul.f32 %f482, %f29, %f30;
|
|
fma.rn.f32 %f22, %f482, %f480, %f22;
|
|
mul.f32 %f483, %f30, %f31;
|
|
fma.rn.f32 %f21, %f483, %f480, %f21;
|
|
fma.rn.f32 %f484, %f31, %f31, 0fBEAAAAAB;
|
|
fma.rn.f32 %f20, %f484, %f480, %f20;
|
|
mul.f32 %f485, %f29, %f31;
|
|
fma.rn.f32 %f19, %f485, %f480, %f19;
|
|
mul.f32 %f486, %f30, %f30;
|
|
mul.f32 %f487, %f29, %f29;
|
|
sub.f32 %f488, %f487, %f486;
|
|
fma.rn.f32 %f18, %f488, %f480, %f18;
|
|
add.f32 %f17, %f17, %f478;
|
|
|
|
BB0_21:
|
|
add.s32 %r454, %r454, 1;
|
|
setp.lt.s32 %p36, %r454, %r2;
|
|
@%p36 bra BB0_6;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_4:
|
|
mov.f32 %f18, %f17;
|
|
mov.f32 %f19, %f17;
|
|
mov.f32 %f20, %f17;
|
|
mov.f32 %f21, %f17;
|
|
mov.f32 %f22, %f17;
|
|
mov.f32 %f23, %f17;
|
|
mov.f32 %f24, %f17;
|
|
mov.f32 %f25, %f17;
|
|
mov.f32 %f26, %f17;
|
|
mov.f32 %f27, %f17;
|
|
|
|
BB0_22:
|
|
mov.pred %p158, %p13;
|
|
|
|
BB0_23:
|
|
cvt.rn.f32.s32 %f489, %r454;
|
|
mov.f32 %f490, 0f3F800000;
|
|
max.f32 %f491, %f489, %f490;
|
|
div.rn.f32 %f1252, %f27, %f491;
|
|
div.rn.f32 %f1262, %f17, %f491;
|
|
div.rn.f32 %f1253, %f26, %f491;
|
|
div.rn.f32 %f1254, %f25, %f491;
|
|
div.rn.f32 %f1255, %f24, %f491;
|
|
div.rn.f32 %f1256, %f23, %f491;
|
|
div.rn.f32 %f1257, %f22, %f491;
|
|
div.rn.f32 %f1258, %f21, %f491;
|
|
div.rn.f32 %f1259, %f20, %f491;
|
|
div.rn.f32 %f1260, %f19, %f491;
|
|
div.rn.f32 %f1261, %f18, %f491;
|
|
@!%p158 bra BB0_76;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_24:
|
|
mov.f32 %f1218, 0f00000000;
|
|
setp.lt.s32 %p37, %r3, 1;
|
|
mov.f32 %f1219, %f1218;
|
|
mov.f32 %f1220, %f1218;
|
|
mov.f32 %f1221, %f1218;
|
|
mov.f32 %f1222, %f1218;
|
|
mov.f32 %f1223, %f1218;
|
|
mov.f32 %f1224, %f1218;
|
|
mov.f32 %f1225, %f1218;
|
|
mov.f32 %f1226, %f1218;
|
|
mov.f32 %f1227, %f1218;
|
|
mov.f32 %f1228, %f1218;
|
|
@%p37 bra BB0_75;
|
|
|
|
cvt.rn.f32.s32 %f514, %r4;
|
|
rcp.rn.f32 %f93, %f514;
|
|
add.u64 %rd68, %SP, 8;
|
|
cvta.to.local.u64 %rd3, %rd68;
|
|
mul.f32 %f94, %f10, 0f3456BF95;
|
|
mul.f32 %f95, %f11, 0f3456BF95;
|
|
mul.f32 %f96, %f12, 0f3456BF95;
|
|
add.u64 %rd69, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd69;
|
|
mov.f32 %f1218, 0f00000000;
|
|
mov.u32 %r154, 0;
|
|
abs.f32 %f515, %f95;
|
|
abs.f32 %f516, %f94;
|
|
max.f32 %f517, %f516, %f515;
|
|
abs.f32 %f518, %f96;
|
|
max.f32 %f519, %f517, %f518;
|
|
mov.u32 %r455, %r154;
|
|
mov.f32 %f1219, %f1218;
|
|
mov.f32 %f1220, %f1218;
|
|
mov.f32 %f1221, %f1218;
|
|
mov.f32 %f1222, %f1218;
|
|
mov.f32 %f1223, %f1218;
|
|
mov.f32 %f1224, %f1218;
|
|
mov.f32 %f1225, %f1218;
|
|
mov.f32 %f1226, %f1218;
|
|
mov.f32 %f1227, %f1218;
|
|
mov.f32 %f1228, %f1218;
|
|
|
|
BB0_26:
|
|
cvt.rn.f32.s32 %f108, %r455;
|
|
max.f32 %f109, %f519, %f335;
|
|
mov.u32 %r457, %r154;
|
|
|
|
BB0_27:
|
|
mad.lo.s32 %r156, %r458, 1664525, 1013904223;
|
|
and.b32 %r157, %r156, 16777215;
|
|
cvt.rn.f32.u32 %f521, %r157;
|
|
fma.rn.f32 %f522, %f521, 0f33800000, %f108;
|
|
mul.f32 %f523, %f93, %f522;
|
|
mad.lo.s32 %r458, %r156, 1664525, 1013904223;
|
|
and.b32 %r158, %r458, 16777215;
|
|
cvt.rn.f32.u32 %f524, %r158;
|
|
cvt.rn.f32.s32 %f525, %r457;
|
|
fma.rn.f32 %f526, %f524, 0f33800000, %f525;
|
|
mul.f32 %f527, %f93, %f526;
|
|
fma.rn.f32 %f121, %f523, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f122, %f121, %f121;
|
|
sub.f32 %f529, %f490, %f122;
|
|
mov.f32 %f530, 0f00000000;
|
|
max.f32 %f531, %f530, %f529;
|
|
sqrt.rn.f32 %f123, %f531;
|
|
mul.f32 %f1235, %f527, 0f40C90FDB;
|
|
abs.f32 %f125, %f1235;
|
|
setp.neu.f32 %p38, %f125, 0f7F800000;
|
|
mov.f32 %f1229, %f1235;
|
|
@%p38 bra BB0_29;
|
|
|
|
mul.rn.f32 %f1229, %f1235, %f530;
|
|
|
|
BB0_29:
|
|
mul.f32 %f533, %f1229, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r468, %f533;
|
|
cvt.rn.f32.s32 %f534, %r468;
|
|
neg.f32 %f535, %f534;
|
|
mov.f32 %f536, 0f3FC90FDA;
|
|
fma.rn.f32 %f537, %f535, %f536, %f1229;
|
|
mov.f32 %f538, 0f33A22168;
|
|
fma.rn.f32 %f539, %f535, %f538, %f537;
|
|
mov.f32 %f540, 0f27C234C5;
|
|
fma.rn.f32 %f1230, %f535, %f540, %f539;
|
|
abs.f32 %f541, %f1229;
|
|
setp.leu.f32 %p39, %f541, 0f47CE4780;
|
|
@%p39 bra BB0_40;
|
|
|
|
mov.b32 %r19, %f1229;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r161, %r19, 8;
|
|
or.b32 %r21, %r161, -2147483648;
|
|
mov.u32 %r459, 0;
|
|
mov.u64 %rd515, 0;
|
|
mov.u64 %rd514, %rd3;
|
|
mov.u32 %r460, %r459;
|
|
|
|
BB0_31:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd71, %rd515, 2;
|
|
mov.u64 %rd72, __cudart_i2opi_f;
|
|
add.s64 %rd73, %rd72, %rd71;
|
|
ld.const.u32 %r164, [%rd73];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r162, %r164, %r21, %r460;
|
|
madc.hi.u32 %r460, %r164, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd514], %r162;
|
|
add.s32 %r459, %r459, 1;
|
|
cvt.s64.s32 %rd515, %r459;
|
|
mul.wide.s32 %rd74, %r459, 4;
|
|
add.s64 %rd514, %rd3, %rd74;
|
|
setp.ne.s32 %p40, %r459, 6;
|
|
@%p40 bra BB0_31;
|
|
|
|
and.b32 %r167, %r20, 255;
|
|
add.s32 %r168, %r167, -128;
|
|
shr.u32 %r169, %r168, 5;
|
|
and.b32 %r26, %r19, -2147483648;
|
|
st.local.u32 [%rd3+24], %r460;
|
|
mov.u32 %r170, 6;
|
|
sub.s32 %r171, %r170, %r169;
|
|
mul.wide.s32 %rd75, %r171, 4;
|
|
add.s64 %rd10, %rd3, %rd75;
|
|
ld.local.u32 %r461, [%rd10];
|
|
ld.local.u32 %r462, [%rd10+-4];
|
|
and.b32 %r29, %r20, 31;
|
|
setp.eq.s32 %p41, %r29, 0;
|
|
@%p41 bra BB0_34;
|
|
|
|
mov.u32 %r172, 32;
|
|
sub.s32 %r173, %r172, %r29;
|
|
shr.u32 %r174, %r462, %r173;
|
|
shl.b32 %r175, %r461, %r29;
|
|
add.s32 %r461, %r174, %r175;
|
|
ld.local.u32 %r176, [%rd10+-8];
|
|
shr.u32 %r177, %r176, %r173;
|
|
shl.b32 %r178, %r462, %r29;
|
|
add.s32 %r462, %r177, %r178;
|
|
|
|
BB0_34:
|
|
shr.u32 %r179, %r462, 30;
|
|
shl.b32 %r180, %r461, 2;
|
|
add.s32 %r463, %r179, %r180;
|
|
shl.b32 %r35, %r462, 2;
|
|
shr.u32 %r181, %r463, 31;
|
|
shr.u32 %r182, %r461, 30;
|
|
add.s32 %r36, %r181, %r182;
|
|
setp.eq.s32 %p42, %r181, 0;
|
|
@%p42 bra BB0_35;
|
|
bra.uni BB0_36;
|
|
|
|
BB0_35:
|
|
mov.u32 %r464, %r26;
|
|
mov.u32 %r465, %r35;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
not.b32 %r183, %r463;
|
|
neg.s32 %r465, %r35;
|
|
setp.eq.s32 %p43, %r35, 0;
|
|
selp.u32 %r184, 1, 0, %p43;
|
|
add.s32 %r463, %r184, %r183;
|
|
xor.b32 %r464, %r26, -2147483648;
|
|
|
|
BB0_37:
|
|
clz.b32 %r467, %r463;
|
|
setp.eq.s32 %p44, %r467, 0;
|
|
shl.b32 %r185, %r463, %r467;
|
|
mov.u32 %r186, 32;
|
|
sub.s32 %r187, %r186, %r467;
|
|
shr.u32 %r188, %r465, %r187;
|
|
add.s32 %r189, %r188, %r185;
|
|
selp.b32 %r44, %r463, %r189, %p44;
|
|
mov.u32 %r190, -921707870;
|
|
mul.hi.u32 %r466, %r44, %r190;
|
|
setp.eq.s32 %p45, %r26, 0;
|
|
neg.s32 %r191, %r36;
|
|
selp.b32 %r468, %r36, %r191, %p45;
|
|
setp.lt.s32 %p46, %r466, 1;
|
|
@%p46 bra BB0_39;
|
|
|
|
mul.lo.s32 %r192, %r44, -921707870;
|
|
shr.u32 %r193, %r192, 31;
|
|
shl.b32 %r194, %r466, 1;
|
|
add.s32 %r466, %r193, %r194;
|
|
add.s32 %r467, %r467, 1;
|
|
|
|
BB0_39:
|
|
mov.u32 %r195, 126;
|
|
sub.s32 %r196, %r195, %r467;
|
|
shl.b32 %r197, %r196, 23;
|
|
add.s32 %r198, %r466, 1;
|
|
shr.u32 %r199, %r198, 7;
|
|
add.s32 %r200, %r199, 1;
|
|
shr.u32 %r201, %r200, 1;
|
|
add.s32 %r202, %r201, %r197;
|
|
or.b32 %r203, %r202, %r464;
|
|
mov.b32 %f1230, %r203;
|
|
|
|
BB0_40:
|
|
mul.rn.f32 %f131, %f1230, %f1230;
|
|
add.s32 %r52, %r468, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p47, %r53, 0;
|
|
@%p47 bra BB0_42;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_42:
|
|
mov.f32 %f544, 0f3C08839E;
|
|
mov.f32 %f545, 0fB94CA1F9;
|
|
fma.rn.f32 %f1231, %f545, %f131, %f544;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_41:
|
|
mov.f32 %f542, 0fBAB6061A;
|
|
mov.f32 %f543, 0f37CCF5CE;
|
|
fma.rn.f32 %f1231, %f543, %f131, %f542;
|
|
|
|
BB0_43:
|
|
@%p47 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f549, 0fBE2AAAA3;
|
|
fma.rn.f32 %f550, %f1231, %f131, %f549;
|
|
fma.rn.f32 %f1232, %f550, %f131, %f530;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f546, 0f3D2AAAA5;
|
|
fma.rn.f32 %f547, %f1231, %f131, %f546;
|
|
mov.f32 %f548, 0fBF000000;
|
|
fma.rn.f32 %f1232, %f547, %f131, %f548;
|
|
|
|
BB0_46:
|
|
fma.rn.f32 %f1233, %f1232, %f1230, %f1230;
|
|
@%p47 bra BB0_48;
|
|
|
|
fma.rn.f32 %f1233, %f1232, %f131, %f490;
|
|
|
|
BB0_48:
|
|
and.b32 %r204, %r52, 2;
|
|
setp.eq.s32 %p50, %r204, 0;
|
|
@%p50 bra BB0_50;
|
|
|
|
mov.f32 %f554, 0fBF800000;
|
|
fma.rn.f32 %f1233, %f1233, %f554, %f530;
|
|
|
|
BB0_50:
|
|
@%p38 bra BB0_52;
|
|
|
|
mul.rn.f32 %f1235, %f1235, %f530;
|
|
|
|
BB0_52:
|
|
mul.f32 %f556, %f1235, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r478, %f556;
|
|
cvt.rn.f32.s32 %f557, %r478;
|
|
neg.f32 %f558, %f557;
|
|
fma.rn.f32 %f560, %f558, %f536, %f1235;
|
|
fma.rn.f32 %f562, %f558, %f538, %f560;
|
|
fma.rn.f32 %f1236, %f558, %f540, %f562;
|
|
abs.f32 %f564, %f1235;
|
|
setp.leu.f32 %p52, %f564, 0f47CE4780;
|
|
@%p52 bra BB0_63;
|
|
|
|
mov.b32 %r55, %f1235;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r207, %r55, 8;
|
|
or.b32 %r57, %r207, -2147483648;
|
|
mov.u32 %r469, 0;
|
|
mov.u64 %rd516, %rd3;
|
|
mov.u64 %rd517, %rd27;
|
|
mov.u32 %r470, %r469;
|
|
|
|
BB0_54:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd77, %rd517, 2;
|
|
mov.u64 %rd78, __cudart_i2opi_f;
|
|
add.s64 %rd79, %rd78, %rd77;
|
|
ld.const.u32 %r210, [%rd79];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r208, %r210, %r57, %r470;
|
|
madc.hi.u32 %r470, %r210, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd516], %r208;
|
|
add.s32 %r469, %r469, 1;
|
|
cvt.s64.s32 %rd517, %r469;
|
|
mul.wide.s32 %rd80, %r469, 4;
|
|
add.s64 %rd516, %rd3, %rd80;
|
|
setp.ne.s32 %p53, %r469, 6;
|
|
@%p53 bra BB0_54;
|
|
|
|
and.b32 %r213, %r56, 255;
|
|
add.s32 %r214, %r213, -128;
|
|
shr.u32 %r215, %r214, 5;
|
|
and.b32 %r62, %r55, -2147483648;
|
|
st.local.u32 [%rd3+24], %r470;
|
|
mov.u32 %r216, 6;
|
|
sub.s32 %r217, %r216, %r215;
|
|
mul.wide.s32 %rd81, %r217, 4;
|
|
add.s64 %rd16, %rd3, %rd81;
|
|
ld.local.u32 %r471, [%rd16];
|
|
ld.local.u32 %r472, [%rd16+-4];
|
|
and.b32 %r65, %r56, 31;
|
|
setp.eq.s32 %p54, %r65, 0;
|
|
@%p54 bra BB0_57;
|
|
|
|
mov.u32 %r218, 32;
|
|
sub.s32 %r219, %r218, %r65;
|
|
shr.u32 %r220, %r472, %r219;
|
|
shl.b32 %r221, %r471, %r65;
|
|
add.s32 %r471, %r220, %r221;
|
|
ld.local.u32 %r222, [%rd16+-8];
|
|
shr.u32 %r223, %r222, %r219;
|
|
shl.b32 %r224, %r472, %r65;
|
|
add.s32 %r472, %r223, %r224;
|
|
|
|
BB0_57:
|
|
shr.u32 %r225, %r472, 30;
|
|
shl.b32 %r226, %r471, 2;
|
|
add.s32 %r473, %r225, %r226;
|
|
shl.b32 %r71, %r472, 2;
|
|
shr.u32 %r227, %r473, 31;
|
|
shr.u32 %r228, %r471, 30;
|
|
add.s32 %r72, %r227, %r228;
|
|
setp.eq.s32 %p55, %r227, 0;
|
|
@%p55 bra BB0_58;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_58:
|
|
mov.u32 %r474, %r62;
|
|
mov.u32 %r475, %r71;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
not.b32 %r229, %r473;
|
|
neg.s32 %r475, %r71;
|
|
setp.eq.s32 %p56, %r71, 0;
|
|
selp.u32 %r230, 1, 0, %p56;
|
|
add.s32 %r473, %r230, %r229;
|
|
xor.b32 %r474, %r62, -2147483648;
|
|
|
|
BB0_60:
|
|
clz.b32 %r477, %r473;
|
|
setp.eq.s32 %p57, %r477, 0;
|
|
shl.b32 %r231, %r473, %r477;
|
|
mov.u32 %r232, 32;
|
|
sub.s32 %r233, %r232, %r477;
|
|
shr.u32 %r234, %r475, %r233;
|
|
add.s32 %r235, %r234, %r231;
|
|
selp.b32 %r80, %r473, %r235, %p57;
|
|
mov.u32 %r236, -921707870;
|
|
mul.hi.u32 %r476, %r80, %r236;
|
|
setp.eq.s32 %p58, %r62, 0;
|
|
neg.s32 %r237, %r72;
|
|
selp.b32 %r478, %r72, %r237, %p58;
|
|
setp.lt.s32 %p59, %r476, 1;
|
|
@%p59 bra BB0_62;
|
|
|
|
mul.lo.s32 %r238, %r80, -921707870;
|
|
shr.u32 %r239, %r238, 31;
|
|
shl.b32 %r240, %r476, 1;
|
|
add.s32 %r476, %r239, %r240;
|
|
add.s32 %r477, %r477, 1;
|
|
|
|
BB0_62:
|
|
mov.u32 %r241, 126;
|
|
sub.s32 %r242, %r241, %r477;
|
|
shl.b32 %r243, %r242, 23;
|
|
add.s32 %r244, %r476, 1;
|
|
shr.u32 %r245, %r244, 7;
|
|
add.s32 %r246, %r245, 1;
|
|
shr.u32 %r247, %r246, 1;
|
|
add.s32 %r248, %r247, %r243;
|
|
or.b32 %r249, %r248, %r474;
|
|
mov.b32 %f1236, %r249;
|
|
|
|
BB0_63:
|
|
mul.rn.f32 %f148, %f1236, %f1236;
|
|
and.b32 %r88, %r478, 1;
|
|
setp.eq.s32 %p60, %r88, 0;
|
|
@%p60 bra BB0_65;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_65:
|
|
mov.f32 %f567, 0f3C08839E;
|
|
mov.f32 %f568, 0fB94CA1F9;
|
|
fma.rn.f32 %f1237, %f568, %f148, %f567;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_64:
|
|
mov.f32 %f565, 0fBAB6061A;
|
|
mov.f32 %f566, 0f37CCF5CE;
|
|
fma.rn.f32 %f1237, %f566, %f148, %f565;
|
|
|
|
BB0_66:
|
|
@%p60 bra BB0_68;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_68:
|
|
mov.f32 %f572, 0fBE2AAAA3;
|
|
fma.rn.f32 %f573, %f1237, %f148, %f572;
|
|
fma.rn.f32 %f1238, %f573, %f148, %f530;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_67:
|
|
mov.f32 %f569, 0f3D2AAAA5;
|
|
fma.rn.f32 %f570, %f1237, %f148, %f569;
|
|
mov.f32 %f571, 0fBF000000;
|
|
fma.rn.f32 %f1238, %f570, %f148, %f571;
|
|
|
|
BB0_69:
|
|
fma.rn.f32 %f1239, %f1238, %f1236, %f1236;
|
|
@%p60 bra BB0_71;
|
|
|
|
fma.rn.f32 %f1239, %f1238, %f148, %f490;
|
|
|
|
BB0_71:
|
|
and.b32 %r250, %r478, 2;
|
|
setp.eq.s32 %p63, %r250, 0;
|
|
@%p63 bra BB0_73;
|
|
|
|
mov.f32 %f577, 0fBF800000;
|
|
fma.rn.f32 %f1239, %f1239, %f577, %f530;
|
|
|
|
BB0_73:
|
|
mul.f32 %f581, %f123, %f1233;
|
|
mov.u32 %r252, 0;
|
|
st.local.u32 [%rd4], %r252;
|
|
ld.global.u32 %r251, [root];
|
|
mul.f32 %f582, %f123, %f1239;
|
|
mov.f32 %f585, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f581, %f582, %f121, %r252, %f109, %f585, %rd69, %r102);
|
|
// inline asm
|
|
ld.local.f32 %f586, [%rd4];
|
|
setp.lt.f32 %p64, %f586, 0f00000000;
|
|
selp.f32 %f587, 0f00000000, %f586, %p64;
|
|
selp.f32 %f588, 0f00000000, 0f3F800000, %p64;
|
|
add.f32 %f1218, %f1218, %f588;
|
|
add.f32 %f1227, %f1227, %f587;
|
|
fma.rn.f32 %f1226, %f582, %f587, %f1226;
|
|
fma.rn.f32 %f1225, %f121, %f587, %f1225;
|
|
fma.rn.f32 %f1224, %f581, %f587, %f1224;
|
|
mul.f32 %f589, %f581, %f582;
|
|
fma.rn.f32 %f1223, %f589, %f587, %f1223;
|
|
mul.f32 %f590, %f121, %f582;
|
|
fma.rn.f32 %f1222, %f590, %f587, %f1222;
|
|
add.f32 %f591, %f122, 0fBEAAAAAB;
|
|
fma.rn.f32 %f1221, %f591, %f587, %f1221;
|
|
mul.f32 %f592, %f121, %f581;
|
|
fma.rn.f32 %f1220, %f592, %f587, %f1220;
|
|
mul.f32 %f593, %f582, %f582;
|
|
mul.f32 %f594, %f581, %f581;
|
|
sub.f32 %f595, %f594, %f593;
|
|
fma.rn.f32 %f1219, %f595, %f587, %f1219;
|
|
mul.f32 %f596, %f1169, %f582;
|
|
fma.rn.f32 %f597, %f1168, %f581, %f596;
|
|
fma.rn.f32 %f598, %f1170, %f121, %f597;
|
|
cvt.sat.f32.f32 %f599, %f598;
|
|
fma.rn.f32 %f1228, %f587, %f599, %f1228;
|
|
add.s32 %r457, %r457, 1;
|
|
setp.lt.s32 %p65, %r457, %r4;
|
|
@%p65 bra BB0_27;
|
|
|
|
add.s32 %r455, %r455, 1;
|
|
setp.lt.s32 %p66, %r455, %r4;
|
|
@%p66 bra BB0_26;
|
|
|
|
BB0_75:
|
|
mul.lo.s32 %r254, %r4, %r4;
|
|
cvt.rn.f32.s32 %f600, %r254;
|
|
div.rn.f32 %f601, %f1228, %f600;
|
|
div.rn.f32 %f1262, %f1218, %f600;
|
|
div.rn.f32 %f1253, %f1227, %f600;
|
|
div.rn.f32 %f1254, %f1226, %f600;
|
|
div.rn.f32 %f1255, %f1225, %f600;
|
|
div.rn.f32 %f1256, %f1224, %f600;
|
|
div.rn.f32 %f1257, %f1223, %f600;
|
|
div.rn.f32 %f1258, %f1222, %f600;
|
|
div.rn.f32 %f1259, %f1221, %f600;
|
|
div.rn.f32 %f1260, %f1220, %f600;
|
|
div.rn.f32 %f1261, %f1219, %f600;
|
|
add.f32 %f1252, %f601, %f601;
|
|
|
|
BB0_76:
|
|
setp.gt.s32 %p67, %r1, -1;
|
|
selp.f32 %f204, %f1262, %f1252, %p67;
|
|
ld.global.u32 %r480, [imageEnabled];
|
|
and.b32 %r255, %r480, 8;
|
|
setp.eq.s32 %p68, %r255, 0;
|
|
@%p68 bra BB0_89;
|
|
|
|
cvt.u64.u32 %rd85, %r6;
|
|
cvt.u64.u32 %rd86, %r7;
|
|
mov.u64 %rd89, image_Mask;
|
|
cvta.global.u64 %rd84, %rd89;
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd84, %r101, %r101, %rd85, %rd86, %rd27, %rd27);
|
|
// inline asm
|
|
mov.f32 %f604, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f605, %f604;
|
|
fma.rn.f32 %f606, %f605, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f205, %f606;
|
|
abs.f32 %f206, %f204;
|
|
setp.lt.f32 %p69, %f206, 0f00800000;
|
|
mul.f32 %f607, %f206, 0f4B800000;
|
|
selp.f32 %f608, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f609, %f607, %f206, %p69;
|
|
mov.b32 %r258, %f609;
|
|
and.b32 %r259, %r258, 8388607;
|
|
or.b32 %r260, %r259, 1065353216;
|
|
mov.b32 %f610, %r260;
|
|
shr.u32 %r261, %r258, 23;
|
|
cvt.rn.f32.u32 %f611, %r261;
|
|
add.f32 %f612, %f608, %f611;
|
|
setp.gt.f32 %p70, %f610, 0f3FB504F3;
|
|
mul.f32 %f613, %f610, 0f3F000000;
|
|
add.f32 %f614, %f612, 0f3F800000;
|
|
selp.f32 %f615, %f613, %f610, %p70;
|
|
selp.f32 %f616, %f614, %f612, %p70;
|
|
add.f32 %f617, %f615, 0fBF800000;
|
|
add.f32 %f603, %f615, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f602,%f603;
|
|
// inline asm
|
|
add.f32 %f618, %f617, %f617;
|
|
mul.f32 %f619, %f602, %f618;
|
|
mul.f32 %f620, %f619, %f619;
|
|
mov.f32 %f621, 0f3C4CAF63;
|
|
mov.f32 %f622, 0f3B18F0FE;
|
|
fma.rn.f32 %f623, %f622, %f620, %f621;
|
|
mov.f32 %f624, 0f3DAAAABD;
|
|
fma.rn.f32 %f625, %f623, %f620, %f624;
|
|
mul.rn.f32 %f626, %f625, %f620;
|
|
mul.rn.f32 %f627, %f626, %f619;
|
|
sub.f32 %f628, %f617, %f619;
|
|
neg.f32 %f629, %f619;
|
|
add.f32 %f630, %f628, %f628;
|
|
fma.rn.f32 %f631, %f629, %f617, %f630;
|
|
mul.rn.f32 %f632, %f602, %f631;
|
|
add.f32 %f633, %f627, %f619;
|
|
sub.f32 %f634, %f619, %f633;
|
|
add.f32 %f635, %f627, %f634;
|
|
add.f32 %f636, %f632, %f635;
|
|
add.f32 %f637, %f633, %f636;
|
|
sub.f32 %f638, %f633, %f637;
|
|
add.f32 %f639, %f636, %f638;
|
|
mov.f32 %f640, 0f3F317200;
|
|
mul.rn.f32 %f641, %f616, %f640;
|
|
mov.f32 %f642, 0f35BFBE8E;
|
|
mul.rn.f32 %f643, %f616, %f642;
|
|
add.f32 %f644, %f641, %f637;
|
|
sub.f32 %f645, %f641, %f644;
|
|
add.f32 %f646, %f637, %f645;
|
|
add.f32 %f647, %f639, %f646;
|
|
add.f32 %f648, %f643, %f647;
|
|
add.f32 %f649, %f644, %f648;
|
|
sub.f32 %f650, %f644, %f649;
|
|
add.f32 %f651, %f648, %f650;
|
|
mov.f32 %f652, 0f3EE8BA2E;
|
|
mul.rn.f32 %f653, %f652, %f649;
|
|
neg.f32 %f654, %f653;
|
|
fma.rn.f32 %f655, %f652, %f649, %f654;
|
|
fma.rn.f32 %f656, %f652, %f651, %f655;
|
|
mov.f32 %f657, 0f00000000;
|
|
fma.rn.f32 %f658, %f657, %f649, %f656;
|
|
add.rn.f32 %f659, %f653, %f658;
|
|
neg.f32 %f660, %f659;
|
|
add.rn.f32 %f661, %f653, %f660;
|
|
add.rn.f32 %f662, %f661, %f658;
|
|
mov.b32 %r262, %f659;
|
|
setp.eq.s32 %p71, %r262, 1118925336;
|
|
add.s32 %r263, %r262, -1;
|
|
mov.b32 %f663, %r263;
|
|
add.f32 %f664, %f662, 0f37000000;
|
|
selp.f32 %f665, %f663, %f659, %p71;
|
|
selp.f32 %f207, %f664, %f662, %p71;
|
|
mul.f32 %f666, %f665, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f667, %f666;
|
|
mov.f32 %f668, 0fBF317200;
|
|
fma.rn.f32 %f669, %f667, %f668, %f665;
|
|
mov.f32 %f670, 0fB5BFBE8E;
|
|
fma.rn.f32 %f671, %f667, %f670, %f669;
|
|
mul.f32 %f672, %f671, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f673, %f672;
|
|
add.f32 %f674, %f667, 0f00000000;
|
|
ex2.approx.f32 %f675, %f674;
|
|
mul.f32 %f676, %f673, %f675;
|
|
setp.lt.f32 %p72, %f665, 0fC2D20000;
|
|
selp.f32 %f677, 0f00000000, %f676, %p72;
|
|
setp.gt.f32 %p73, %f665, 0f42D20000;
|
|
selp.f32 %f1263, 0f7F800000, %f677, %p73;
|
|
setp.eq.f32 %p74, %f1263, 0f7F800000;
|
|
@%p74 bra BB0_79;
|
|
|
|
fma.rn.f32 %f1263, %f1263, %f207, %f1263;
|
|
|
|
BB0_79:
|
|
setp.lt.f32 %p75, %f204, 0f00000000;
|
|
setp.eq.f32 %p76, %f205, 0f3F800000;
|
|
and.pred %p3, %p75, %p76;
|
|
mov.b32 %r264, %f1263;
|
|
xor.b32 %r265, %r264, -2147483648;
|
|
mov.b32 %f678, %r265;
|
|
selp.f32 %f1265, %f678, %f1263, %p3;
|
|
setp.eq.f32 %p77, %f204, 0f00000000;
|
|
@%p77 bra BB0_82;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_82:
|
|
add.f32 %f681, %f204, %f204;
|
|
selp.f32 %f1265, %f681, 0f00000000, %p76;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_162:
|
|
mov.u64 %rd330, image_HDR;
|
|
cvta.global.u64 %rd325, %rd330;
|
|
mov.u32 %r386, 8;
|
|
mov.u64 %rd329, 0;
|
|
// inline asm
|
|
call (%rd324), _rt_buffer_get_64, (%rd325, %r101, %r386, %rd20, %rd21, %rd329, %rd329);
|
|
// inline asm
|
|
mov.f32 %f1097, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f1097;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs181, 0;
|
|
st.v4.u16 [%rd324], {%rs180, %rs180, %rs180, %rs181};
|
|
|
|
BB0_163:
|
|
ld.global.u32 %r387, [additive];
|
|
setp.eq.s32 %p149, %r387, 0;
|
|
@%p149 bra BB0_165;
|
|
|
|
mov.u64 %rd343, image_RNM0;
|
|
cvta.global.u64 %rd332, %rd343;
|
|
mov.u32 %r391, 8;
|
|
mov.u64 %rd342, 0;
|
|
// inline asm
|
|
call (%rd331), _rt_buffer_get_64, (%rd332, %r101, %r391, %rd20, %rd21, %rd342, %rd342);
|
|
// inline asm
|
|
ld.v4.u16 {%rs188, %rs189, %rs190, %rs191}, [%rd331];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1098, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1099, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1100, %rs190;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd337), _rt_buffer_get_64, (%rd332, %r101, %r391, %rd20, %rd21, %rd342, %rd342);
|
|
// inline asm
|
|
add.f32 %f1101, %f1098, 0f00000000;
|
|
add.f32 %f1102, %f1099, 0f00000000;
|
|
add.f32 %f1103, %f1100, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs187, %f1103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f1102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f1101;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs192, 0;
|
|
st.v4.u16 [%rd337], {%rs185, %rs186, %rs187, %rs192};
|
|
bra.uni BB0_166;
|
|
|
|
BB0_165:
|
|
mov.u64 %rd350, image_RNM0;
|
|
cvta.global.u64 %rd345, %rd350;
|
|
mov.u32 %r393, 8;
|
|
mov.u64 %rd349, 0;
|
|
// inline asm
|
|
call (%rd344), _rt_buffer_get_64, (%rd345, %r101, %r393, %rd20, %rd21, %rd349, %rd349);
|
|
// inline asm
|
|
mov.f32 %f1104, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs193, %f1104;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs194, 0;
|
|
st.v4.u16 [%rd344], {%rs193, %rs193, %rs193, %rs194};
|
|
|
|
BB0_166:
|
|
ld.global.u32 %r394, [additive];
|
|
setp.eq.s32 %p150, %r394, 0;
|
|
@%p150 bra BB0_168;
|
|
|
|
mov.u64 %rd363, image_RNM1;
|
|
cvta.global.u64 %rd352, %rd363;
|
|
mov.u32 %r398, 8;
|
|
mov.u64 %rd362, 0;
|
|
// inline asm
|
|
call (%rd351), _rt_buffer_get_64, (%rd352, %r101, %r398, %rd20, %rd21, %rd362, %rd362);
|
|
// inline asm
|
|
ld.v4.u16 {%rs201, %rs202, %rs203, %rs204}, [%rd351];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1105, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1106, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1107, %rs203;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd357), _rt_buffer_get_64, (%rd352, %r101, %r398, %rd20, %rd21, %rd362, %rd362);
|
|
// inline asm
|
|
add.f32 %f1108, %f1105, 0f00000000;
|
|
add.f32 %f1109, %f1106, 0f00000000;
|
|
add.f32 %f1110, %f1107, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs200, %f1110;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f1109;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f1108;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs205, 0;
|
|
st.v4.u16 [%rd357], {%rs198, %rs199, %rs200, %rs205};
|
|
bra.uni BB0_169;
|
|
|
|
BB0_168:
|
|
mov.u64 %rd370, image_RNM1;
|
|
cvta.global.u64 %rd365, %rd370;
|
|
mov.u32 %r400, 8;
|
|
mov.u64 %rd369, 0;
|
|
// inline asm
|
|
call (%rd364), _rt_buffer_get_64, (%rd365, %r101, %r400, %rd20, %rd21, %rd369, %rd369);
|
|
// inline asm
|
|
mov.f32 %f1111, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f1111;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs207, 0;
|
|
st.v4.u16 [%rd364], {%rs206, %rs206, %rs206, %rs207};
|
|
|
|
BB0_169:
|
|
ld.global.u32 %r401, [additive];
|
|
setp.eq.s32 %p151, %r401, 0;
|
|
@%p151 bra BB0_171;
|
|
|
|
mov.u64 %rd383, image_RNM2;
|
|
cvta.global.u64 %rd372, %rd383;
|
|
mov.u32 %r405, 8;
|
|
mov.u64 %rd382, 0;
|
|
// inline asm
|
|
call (%rd371), _rt_buffer_get_64, (%rd372, %r101, %r405, %rd20, %rd21, %rd382, %rd382);
|
|
// inline asm
|
|
ld.v4.u16 {%rs214, %rs215, %rs216, %rs217}, [%rd371];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1112, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1113, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1114, %rs216;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd377), _rt_buffer_get_64, (%rd372, %r101, %r405, %rd20, %rd21, %rd382, %rd382);
|
|
// inline asm
|
|
add.f32 %f1115, %f1112, 0f00000000;
|
|
add.f32 %f1116, %f1113, 0f00000000;
|
|
add.f32 %f1117, %f1114, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs213, %f1117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f1116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f1115;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs218, 0;
|
|
st.v4.u16 [%rd377], {%rs211, %rs212, %rs213, %rs218};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_171:
|
|
mov.u64 %rd390, image_RNM2;
|
|
cvta.global.u64 %rd385, %rd390;
|
|
mov.u32 %r407, 8;
|
|
mov.u64 %rd389, 0;
|
|
// inline asm
|
|
call (%rd384), _rt_buffer_get_64, (%rd385, %r101, %r407, %rd20, %rd21, %rd389, %rd389);
|
|
// inline asm
|
|
mov.f32 %f1118, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs219, %f1118;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs220, 0;
|
|
st.v4.u16 [%rd384], {%rs219, %rs219, %rs219, %rs220};
|
|
|
|
BB0_172:
|
|
ld.global.u32 %r408, [additive];
|
|
setp.eq.s32 %p152, %r408, 0;
|
|
@%p152 bra BB0_174;
|
|
|
|
mov.u64 %rd403, image_RNM3;
|
|
cvta.global.u64 %rd392, %rd403;
|
|
mov.u32 %r412, 8;
|
|
mov.u64 %rd402, 0;
|
|
// inline asm
|
|
call (%rd391), _rt_buffer_get_64, (%rd392, %r101, %r412, %rd20, %rd21, %rd402, %rd402);
|
|
// inline asm
|
|
ld.v4.u16 {%rs227, %rs228, %rs229, %rs230}, [%rd391];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1119, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1120, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1121, %rs229;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd397), _rt_buffer_get_64, (%rd392, %r101, %r412, %rd20, %rd21, %rd402, %rd402);
|
|
// inline asm
|
|
add.f32 %f1122, %f1119, 0f00000000;
|
|
add.f32 %f1123, %f1120, 0f00000000;
|
|
add.f32 %f1124, %f1121, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs226, %f1124;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f1123;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f1122;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs231, 0;
|
|
st.v4.u16 [%rd397], {%rs224, %rs225, %rs226, %rs231};
|
|
bra.uni BB0_175;
|
|
|
|
BB0_174:
|
|
mov.u64 %rd410, image_RNM3;
|
|
cvta.global.u64 %rd405, %rd410;
|
|
mov.u32 %r414, 8;
|
|
mov.u64 %rd409, 0;
|
|
// inline asm
|
|
call (%rd404), _rt_buffer_get_64, (%rd405, %r101, %r414, %rd20, %rd21, %rd409, %rd409);
|
|
// inline asm
|
|
mov.f32 %f1125, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs232, %f1125;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs233, 0;
|
|
st.v4.u16 [%rd404], {%rs232, %rs232, %rs232, %rs233};
|
|
|
|
BB0_175:
|
|
ld.global.u32 %r415, [additive];
|
|
setp.eq.s32 %p153, %r415, 0;
|
|
@%p153 bra BB0_177;
|
|
|
|
mov.u64 %rd423, image_RNM4;
|
|
cvta.global.u64 %rd412, %rd423;
|
|
mov.u32 %r419, 8;
|
|
mov.u64 %rd422, 0;
|
|
// inline asm
|
|
call (%rd411), _rt_buffer_get_64, (%rd412, %r101, %r419, %rd20, %rd21, %rd422, %rd422);
|
|
// inline asm
|
|
ld.v4.u16 {%rs240, %rs241, %rs242, %rs243}, [%rd411];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1126, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1127, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1128, %rs242;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd417), _rt_buffer_get_64, (%rd412, %r101, %r419, %rd20, %rd21, %rd422, %rd422);
|
|
// inline asm
|
|
add.f32 %f1129, %f1126, 0f00000000;
|
|
add.f32 %f1130, %f1127, 0f00000000;
|
|
add.f32 %f1131, %f1128, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs239, %f1131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f1130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f1129;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs244, 0;
|
|
st.v4.u16 [%rd417], {%rs237, %rs238, %rs239, %rs244};
|
|
bra.uni BB0_178;
|
|
|
|
BB0_177:
|
|
mov.u64 %rd430, image_RNM4;
|
|
cvta.global.u64 %rd425, %rd430;
|
|
mov.u32 %r421, 8;
|
|
mov.u64 %rd429, 0;
|
|
// inline asm
|
|
call (%rd424), _rt_buffer_get_64, (%rd425, %r101, %r421, %rd20, %rd21, %rd429, %rd429);
|
|
// inline asm
|
|
mov.f32 %f1132, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs245, %f1132;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs246, 0;
|
|
st.v4.u16 [%rd424], {%rs245, %rs245, %rs245, %rs246};
|
|
|
|
BB0_178:
|
|
ld.global.u32 %r422, [additive];
|
|
setp.eq.s32 %p154, %r422, 0;
|
|
@%p154 bra BB0_180;
|
|
|
|
mov.u64 %rd443, image_RNM5;
|
|
cvta.global.u64 %rd432, %rd443;
|
|
mov.u32 %r426, 8;
|
|
mov.u64 %rd442, 0;
|
|
// inline asm
|
|
call (%rd431), _rt_buffer_get_64, (%rd432, %r101, %r426, %rd20, %rd21, %rd442, %rd442);
|
|
// inline asm
|
|
ld.v4.u16 {%rs253, %rs254, %rs255, %rs256}, [%rd431];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1133, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1134, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1135, %rs255;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd437), _rt_buffer_get_64, (%rd432, %r101, %r426, %rd20, %rd21, %rd442, %rd442);
|
|
// inline asm
|
|
add.f32 %f1136, %f1133, 0f00000000;
|
|
add.f32 %f1137, %f1134, 0f00000000;
|
|
add.f32 %f1138, %f1135, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs252, %f1138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f1137;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f1136;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs257, 0;
|
|
st.v4.u16 [%rd437], {%rs250, %rs251, %rs252, %rs257};
|
|
bra.uni BB0_181;
|
|
|
|
BB0_180:
|
|
mov.u64 %rd450, image_RNM5;
|
|
cvta.global.u64 %rd445, %rd450;
|
|
mov.u32 %r428, 8;
|
|
mov.u64 %rd449, 0;
|
|
// inline asm
|
|
call (%rd444), _rt_buffer_get_64, (%rd445, %r101, %r428, %rd20, %rd21, %rd449, %rd449);
|
|
// inline asm
|
|
mov.f32 %f1139, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs258, %f1139;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs259, 0;
|
|
st.v4.u16 [%rd444], {%rs258, %rs258, %rs258, %rs259};
|
|
|
|
BB0_181:
|
|
ld.global.u32 %r429, [additive];
|
|
setp.eq.s32 %p155, %r429, 0;
|
|
@%p155 bra BB0_183;
|
|
|
|
mov.u64 %rd463, image_RNM6;
|
|
cvta.global.u64 %rd452, %rd463;
|
|
mov.u32 %r433, 8;
|
|
mov.u64 %rd462, 0;
|
|
// inline asm
|
|
call (%rd451), _rt_buffer_get_64, (%rd452, %r101, %r433, %rd20, %rd21, %rd462, %rd462);
|
|
// inline asm
|
|
ld.v4.u16 {%rs266, %rs267, %rs268, %rs269}, [%rd451];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1140, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1141, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1142, %rs268;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd457), _rt_buffer_get_64, (%rd452, %r101, %r433, %rd20, %rd21, %rd462, %rd462);
|
|
// inline asm
|
|
add.f32 %f1143, %f1140, 0f00000000;
|
|
add.f32 %f1144, %f1141, 0f00000000;
|
|
add.f32 %f1145, %f1142, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs265, %f1145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f1144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f1143;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs270, 0;
|
|
st.v4.u16 [%rd457], {%rs263, %rs264, %rs265, %rs270};
|
|
bra.uni BB0_184;
|
|
|
|
BB0_183:
|
|
mov.u64 %rd470, image_RNM6;
|
|
cvta.global.u64 %rd465, %rd470;
|
|
mov.u32 %r435, 8;
|
|
mov.u64 %rd469, 0;
|
|
// inline asm
|
|
call (%rd464), _rt_buffer_get_64, (%rd465, %r101, %r435, %rd20, %rd21, %rd469, %rd469);
|
|
// inline asm
|
|
mov.f32 %f1146, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs271, %f1146;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs272, 0;
|
|
st.v4.u16 [%rd464], {%rs271, %rs271, %rs271, %rs272};
|
|
|
|
BB0_184:
|
|
ld.global.u32 %r436, [additive];
|
|
setp.eq.s32 %p156, %r436, 0;
|
|
@%p156 bra BB0_186;
|
|
|
|
mov.u64 %rd483, image_RNM7;
|
|
cvta.global.u64 %rd472, %rd483;
|
|
mov.u32 %r440, 8;
|
|
mov.u64 %rd482, 0;
|
|
// inline asm
|
|
call (%rd471), _rt_buffer_get_64, (%rd472, %r101, %r440, %rd20, %rd21, %rd482, %rd482);
|
|
// inline asm
|
|
ld.v4.u16 {%rs279, %rs280, %rs281, %rs282}, [%rd471];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1147, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1148, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1149, %rs281;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd477), _rt_buffer_get_64, (%rd472, %r101, %r440, %rd20, %rd21, %rd482, %rd482);
|
|
// inline asm
|
|
add.f32 %f1150, %f1147, 0f00000000;
|
|
add.f32 %f1151, %f1148, 0f00000000;
|
|
add.f32 %f1152, %f1149, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs278, %f1152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f1151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f1150;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs283, 0;
|
|
st.v4.u16 [%rd477], {%rs276, %rs277, %rs278, %rs283};
|
|
bra.uni BB0_187;
|
|
|
|
BB0_186:
|
|
mov.u64 %rd490, image_RNM7;
|
|
cvta.global.u64 %rd485, %rd490;
|
|
mov.u32 %r442, 8;
|
|
mov.u64 %rd489, 0;
|
|
// inline asm
|
|
call (%rd484), _rt_buffer_get_64, (%rd485, %r101, %r442, %rd20, %rd21, %rd489, %rd489);
|
|
// inline asm
|
|
mov.f32 %f1153, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs284, %f1153;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs285, 0;
|
|
st.v4.u16 [%rd484], {%rs284, %rs284, %rs284, %rs285};
|
|
|
|
BB0_187:
|
|
ld.global.u32 %r443, [additive];
|
|
setp.eq.s32 %p157, %r443, 0;
|
|
@%p157 bra BB0_189;
|
|
|
|
mov.u64 %rd503, image_RNM8;
|
|
cvta.global.u64 %rd492, %rd503;
|
|
mov.u32 %r447, 8;
|
|
mov.u64 %rd502, 0;
|
|
// inline asm
|
|
call (%rd491), _rt_buffer_get_64, (%rd492, %r101, %r447, %rd20, %rd21, %rd502, %rd502);
|
|
// inline asm
|
|
ld.v4.u16 {%rs292, %rs293, %rs294, %rs295}, [%rd491];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1154, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1155, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1156, %rs294;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd497), _rt_buffer_get_64, (%rd492, %r101, %r447, %rd20, %rd21, %rd502, %rd502);
|
|
// inline asm
|
|
add.f32 %f1157, %f1154, 0f00000000;
|
|
add.f32 %f1158, %f1155, 0f00000000;
|
|
add.f32 %f1159, %f1156, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs291, %f1159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f1158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f1157;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs296, 0;
|
|
st.v4.u16 [%rd497], {%rs289, %rs290, %rs291, %rs296};
|
|
bra.uni BB0_190;
|
|
|
|
BB0_189:
|
|
mov.u64 %rd510, image_RNM8;
|
|
cvta.global.u64 %rd505, %rd510;
|
|
mov.u32 %r449, 8;
|
|
mov.u64 %rd509, 0;
|
|
// inline asm
|
|
call (%rd504), _rt_buffer_get_64, (%rd505, %r101, %r449, %rd20, %rd21, %rd509, %rd509);
|
|
// inline asm
|
|
mov.f32 %f1160, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs297, %f1160;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs298, 0;
|
|
st.v4.u16 [%rd504], {%rs297, %rs297, %rs297, %rs298};
|
|
bra.uni BB0_190;
|
|
|
|
BB0_80:
|
|
setp.geu.f32 %p78, %f204, 0f00000000;
|
|
@%p78 bra BB0_83;
|
|
|
|
cvt.rzi.f32.f32 %f680, %f652;
|
|
setp.neu.f32 %p79, %f680, 0f3EE8BA2E;
|
|
selp.f32 %f1265, 0f7FFFFFFF, %f1265, %p79;
|
|
|
|
BB0_83:
|
|
add.f32 %f682, %f206, 0f3EE8BA2E;
|
|
mov.b32 %r266, %f682;
|
|
setp.lt.s32 %p81, %r266, 2139095040;
|
|
@%p81 bra BB0_88;
|
|
|
|
setp.gtu.f32 %p82, %f206, 0f7F800000;
|
|
@%p82 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f1265, %f204, 0f3EE8BA2E;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
setp.neu.f32 %p83, %f206, 0f7F800000;
|
|
@%p83 bra BB0_88;
|
|
|
|
selp.f32 %f1265, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_88:
|
|
mul.f32 %f683, %f1265, 0f437F0000;
|
|
setp.eq.f32 %p84, %f204, 0f3F800000;
|
|
selp.f32 %f684, 0f437F0000, %f683, %p84;
|
|
cvt.rzi.u32.f32 %r267, %f684;
|
|
cvt.u16.u32 %rs20, %r267;
|
|
mov.u16 %rs21, 255;
|
|
st.v2.u8 [%rd83], {%rs20, %rs21};
|
|
ld.global.u32 %r480, [imageEnabled];
|
|
|
|
BB0_89:
|
|
ld.global.f32 %f685, [lightColor];
|
|
mul.f32 %f218, %f1252, %f685;
|
|
ld.global.f32 %f686, [lightColor+4];
|
|
mul.f32 %f219, %f1252, %f686;
|
|
ld.global.f32 %f687, [lightColor+8];
|
|
mul.f32 %f220, %f1252, %f687;
|
|
and.b32 %r268, %r480, 1;
|
|
setp.eq.b32 %p85, %r268, 1;
|
|
@!%p85 bra BB0_124;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_90:
|
|
mov.f32 %f690, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f691, %f690;
|
|
fma.rn.f32 %f692, %f691, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f221, %f692;
|
|
abs.f32 %f222, %f218;
|
|
setp.lt.f32 %p86, %f222, 0f00800000;
|
|
mul.f32 %f693, %f222, 0f4B800000;
|
|
selp.f32 %f694, 0fC3170000, 0fC2FE0000, %p86;
|
|
selp.f32 %f695, %f693, %f222, %p86;
|
|
mov.b32 %r269, %f695;
|
|
and.b32 %r270, %r269, 8388607;
|
|
or.b32 %r271, %r270, 1065353216;
|
|
mov.b32 %f696, %r271;
|
|
shr.u32 %r272, %r269, 23;
|
|
cvt.rn.f32.u32 %f697, %r272;
|
|
add.f32 %f698, %f694, %f697;
|
|
setp.gt.f32 %p87, %f696, 0f3FB504F3;
|
|
mul.f32 %f699, %f696, 0f3F000000;
|
|
add.f32 %f700, %f698, 0f3F800000;
|
|
selp.f32 %f701, %f699, %f696, %p87;
|
|
selp.f32 %f702, %f700, %f698, %p87;
|
|
add.f32 %f703, %f701, 0fBF800000;
|
|
add.f32 %f689, %f701, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f688,%f689;
|
|
// inline asm
|
|
add.f32 %f704, %f703, %f703;
|
|
mul.f32 %f705, %f688, %f704;
|
|
mul.f32 %f706, %f705, %f705;
|
|
mov.f32 %f707, 0f3C4CAF63;
|
|
mov.f32 %f708, 0f3B18F0FE;
|
|
fma.rn.f32 %f709, %f708, %f706, %f707;
|
|
mov.f32 %f710, 0f3DAAAABD;
|
|
fma.rn.f32 %f711, %f709, %f706, %f710;
|
|
mul.rn.f32 %f712, %f711, %f706;
|
|
mul.rn.f32 %f713, %f712, %f705;
|
|
sub.f32 %f714, %f703, %f705;
|
|
neg.f32 %f715, %f705;
|
|
add.f32 %f716, %f714, %f714;
|
|
fma.rn.f32 %f717, %f715, %f703, %f716;
|
|
mul.rn.f32 %f718, %f688, %f717;
|
|
add.f32 %f719, %f713, %f705;
|
|
sub.f32 %f720, %f705, %f719;
|
|
add.f32 %f721, %f713, %f720;
|
|
add.f32 %f722, %f718, %f721;
|
|
add.f32 %f723, %f719, %f722;
|
|
sub.f32 %f724, %f719, %f723;
|
|
add.f32 %f725, %f722, %f724;
|
|
mov.f32 %f726, 0f3F317200;
|
|
mul.rn.f32 %f727, %f702, %f726;
|
|
mov.f32 %f728, 0f35BFBE8E;
|
|
mul.rn.f32 %f729, %f702, %f728;
|
|
add.f32 %f730, %f727, %f723;
|
|
sub.f32 %f731, %f727, %f730;
|
|
add.f32 %f732, %f723, %f731;
|
|
add.f32 %f733, %f725, %f732;
|
|
add.f32 %f734, %f729, %f733;
|
|
add.f32 %f735, %f730, %f734;
|
|
sub.f32 %f736, %f730, %f735;
|
|
add.f32 %f737, %f734, %f736;
|
|
mov.f32 %f738, 0f3EE66666;
|
|
mul.rn.f32 %f739, %f738, %f735;
|
|
neg.f32 %f740, %f739;
|
|
fma.rn.f32 %f741, %f738, %f735, %f740;
|
|
fma.rn.f32 %f742, %f738, %f737, %f741;
|
|
mov.f32 %f743, 0f00000000;
|
|
fma.rn.f32 %f744, %f743, %f735, %f742;
|
|
add.rn.f32 %f745, %f739, %f744;
|
|
neg.f32 %f746, %f745;
|
|
add.rn.f32 %f747, %f739, %f746;
|
|
add.rn.f32 %f748, %f747, %f744;
|
|
mov.b32 %r273, %f745;
|
|
setp.eq.s32 %p88, %r273, 1118925336;
|
|
add.s32 %r274, %r273, -1;
|
|
mov.b32 %f749, %r274;
|
|
add.f32 %f750, %f748, 0f37000000;
|
|
selp.f32 %f751, %f749, %f745, %p88;
|
|
selp.f32 %f223, %f750, %f748, %p88;
|
|
mul.f32 %f752, %f751, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f753, %f752;
|
|
mov.f32 %f754, 0fBF317200;
|
|
fma.rn.f32 %f755, %f753, %f754, %f751;
|
|
mov.f32 %f756, 0fB5BFBE8E;
|
|
fma.rn.f32 %f757, %f753, %f756, %f755;
|
|
mul.f32 %f758, %f757, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f759, %f758;
|
|
add.f32 %f760, %f753, 0f00000000;
|
|
ex2.approx.f32 %f761, %f760;
|
|
mul.f32 %f762, %f759, %f761;
|
|
setp.lt.f32 %p89, %f751, 0fC2D20000;
|
|
selp.f32 %f763, 0f00000000, %f762, %p89;
|
|
setp.gt.f32 %p90, %f751, 0f42D20000;
|
|
selp.f32 %f1266, 0f7F800000, %f763, %p90;
|
|
setp.eq.f32 %p91, %f1266, 0f7F800000;
|
|
@%p91 bra BB0_92;
|
|
|
|
fma.rn.f32 %f1266, %f1266, %f223, %f1266;
|
|
|
|
BB0_92:
|
|
setp.lt.f32 %p92, %f218, 0f00000000;
|
|
setp.eq.f32 %p93, %f221, 0f3F800000;
|
|
and.pred %p4, %p92, %p93;
|
|
mov.b32 %r275, %f1266;
|
|
xor.b32 %r276, %r275, -2147483648;
|
|
mov.b32 %f764, %r276;
|
|
selp.f32 %f1268, %f764, %f1266, %p4;
|
|
setp.eq.f32 %p94, %f218, 0f00000000;
|
|
@%p94 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f767, %f218, %f218;
|
|
selp.f32 %f1268, %f767, 0f00000000, %p93;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.geu.f32 %p95, %f218, 0f00000000;
|
|
@%p95 bra BB0_96;
|
|
|
|
cvt.rzi.f32.f32 %f766, %f738;
|
|
setp.neu.f32 %p96, %f766, 0f3EE66666;
|
|
selp.f32 %f1268, 0f7FFFFFFF, %f1268, %p96;
|
|
|
|
BB0_96:
|
|
add.f32 %f768, %f222, 0f3EE66666;
|
|
mov.b32 %r277, %f768;
|
|
setp.lt.s32 %p98, %r277, 2139095040;
|
|
@%p98 bra BB0_101;
|
|
|
|
setp.gtu.f32 %p99, %f222, 0f7F800000;
|
|
@%p99 bra BB0_100;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_100:
|
|
add.f32 %f1268, %f218, 0f3EE66666;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_98:
|
|
setp.neu.f32 %p100, %f222, 0f7F800000;
|
|
@%p100 bra BB0_101;
|
|
|
|
selp.f32 %f1268, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_101:
|
|
setp.eq.f32 %p101, %f218, 0f3F800000;
|
|
selp.f32 %f234, 0f3F800000, %f1268, %p101;
|
|
abs.f32 %f235, %f219;
|
|
setp.lt.f32 %p102, %f235, 0f00800000;
|
|
mul.f32 %f771, %f235, 0f4B800000;
|
|
selp.f32 %f772, 0fC3170000, 0fC2FE0000, %p102;
|
|
selp.f32 %f773, %f771, %f235, %p102;
|
|
mov.b32 %r278, %f773;
|
|
and.b32 %r279, %r278, 8388607;
|
|
or.b32 %r280, %r279, 1065353216;
|
|
mov.b32 %f774, %r280;
|
|
shr.u32 %r281, %r278, 23;
|
|
cvt.rn.f32.u32 %f775, %r281;
|
|
add.f32 %f776, %f772, %f775;
|
|
setp.gt.f32 %p103, %f774, 0f3FB504F3;
|
|
mul.f32 %f777, %f774, 0f3F000000;
|
|
add.f32 %f778, %f776, 0f3F800000;
|
|
selp.f32 %f779, %f777, %f774, %p103;
|
|
selp.f32 %f780, %f778, %f776, %p103;
|
|
add.f32 %f781, %f779, 0fBF800000;
|
|
add.f32 %f770, %f779, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f769,%f770;
|
|
// inline asm
|
|
add.f32 %f782, %f781, %f781;
|
|
mul.f32 %f783, %f769, %f782;
|
|
mul.f32 %f784, %f783, %f783;
|
|
fma.rn.f32 %f787, %f708, %f784, %f707;
|
|
fma.rn.f32 %f789, %f787, %f784, %f710;
|
|
mul.rn.f32 %f790, %f789, %f784;
|
|
mul.rn.f32 %f791, %f790, %f783;
|
|
sub.f32 %f792, %f781, %f783;
|
|
neg.f32 %f793, %f783;
|
|
add.f32 %f794, %f792, %f792;
|
|
fma.rn.f32 %f795, %f793, %f781, %f794;
|
|
mul.rn.f32 %f796, %f769, %f795;
|
|
add.f32 %f797, %f791, %f783;
|
|
sub.f32 %f798, %f783, %f797;
|
|
add.f32 %f799, %f791, %f798;
|
|
add.f32 %f800, %f796, %f799;
|
|
add.f32 %f801, %f797, %f800;
|
|
sub.f32 %f802, %f797, %f801;
|
|
add.f32 %f803, %f800, %f802;
|
|
mul.rn.f32 %f805, %f780, %f726;
|
|
mul.rn.f32 %f807, %f780, %f728;
|
|
add.f32 %f808, %f805, %f801;
|
|
sub.f32 %f809, %f805, %f808;
|
|
add.f32 %f810, %f801, %f809;
|
|
add.f32 %f811, %f803, %f810;
|
|
add.f32 %f812, %f807, %f811;
|
|
add.f32 %f813, %f808, %f812;
|
|
sub.f32 %f814, %f808, %f813;
|
|
add.f32 %f815, %f812, %f814;
|
|
mul.rn.f32 %f817, %f738, %f813;
|
|
neg.f32 %f818, %f817;
|
|
fma.rn.f32 %f819, %f738, %f813, %f818;
|
|
fma.rn.f32 %f820, %f738, %f815, %f819;
|
|
fma.rn.f32 %f822, %f743, %f813, %f820;
|
|
add.rn.f32 %f823, %f817, %f822;
|
|
neg.f32 %f824, %f823;
|
|
add.rn.f32 %f825, %f817, %f824;
|
|
add.rn.f32 %f826, %f825, %f822;
|
|
mov.b32 %r282, %f823;
|
|
setp.eq.s32 %p104, %r282, 1118925336;
|
|
add.s32 %r283, %r282, -1;
|
|
mov.b32 %f827, %r283;
|
|
add.f32 %f828, %f826, 0f37000000;
|
|
selp.f32 %f829, %f827, %f823, %p104;
|
|
selp.f32 %f236, %f828, %f826, %p104;
|
|
mul.f32 %f830, %f829, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f831, %f830;
|
|
fma.rn.f32 %f833, %f831, %f754, %f829;
|
|
fma.rn.f32 %f835, %f831, %f756, %f833;
|
|
mul.f32 %f836, %f835, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f837, %f836;
|
|
add.f32 %f838, %f831, 0f00000000;
|
|
ex2.approx.f32 %f839, %f838;
|
|
mul.f32 %f840, %f837, %f839;
|
|
setp.lt.f32 %p105, %f829, 0fC2D20000;
|
|
selp.f32 %f841, 0f00000000, %f840, %p105;
|
|
setp.gt.f32 %p106, %f829, 0f42D20000;
|
|
selp.f32 %f1269, 0f7F800000, %f841, %p106;
|
|
setp.eq.f32 %p107, %f1269, 0f7F800000;
|
|
@%p107 bra BB0_103;
|
|
|
|
fma.rn.f32 %f1269, %f1269, %f236, %f1269;
|
|
|
|
BB0_103:
|
|
setp.lt.f32 %p108, %f219, 0f00000000;
|
|
and.pred %p5, %p108, %p93;
|
|
mov.b32 %r284, %f1269;
|
|
xor.b32 %r285, %r284, -2147483648;
|
|
mov.b32 %f842, %r285;
|
|
selp.f32 %f1271, %f842, %f1269, %p5;
|
|
setp.eq.f32 %p110, %f219, 0f00000000;
|
|
@%p110 bra BB0_106;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_106:
|
|
add.f32 %f845, %f219, %f219;
|
|
selp.f32 %f1271, %f845, 0f00000000, %p93;
|
|
bra.uni BB0_107;
|
|
|
|
BB0_104:
|
|
setp.geu.f32 %p111, %f219, 0f00000000;
|
|
@%p111 bra BB0_107;
|
|
|
|
cvt.rzi.f32.f32 %f844, %f738;
|
|
setp.neu.f32 %p112, %f844, 0f3EE66666;
|
|
selp.f32 %f1271, 0f7FFFFFFF, %f1271, %p112;
|
|
|
|
BB0_107:
|
|
add.f32 %f846, %f235, 0f3EE66666;
|
|
mov.b32 %r286, %f846;
|
|
setp.lt.s32 %p114, %r286, 2139095040;
|
|
@%p114 bra BB0_112;
|
|
|
|
setp.gtu.f32 %p115, %f235, 0f7F800000;
|
|
@%p115 bra BB0_111;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_111:
|
|
add.f32 %f1271, %f219, 0f3EE66666;
|
|
bra.uni BB0_112;
|
|
|
|
BB0_109:
|
|
setp.neu.f32 %p116, %f235, 0f7F800000;
|
|
@%p116 bra BB0_112;
|
|
|
|
selp.f32 %f1271, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_112:
|
|
setp.eq.f32 %p117, %f219, 0f3F800000;
|
|
selp.f32 %f247, 0f3F800000, %f1271, %p117;
|
|
abs.f32 %f248, %f220;
|
|
setp.lt.f32 %p118, %f248, 0f00800000;
|
|
mul.f32 %f849, %f248, 0f4B800000;
|
|
selp.f32 %f850, 0fC3170000, 0fC2FE0000, %p118;
|
|
selp.f32 %f851, %f849, %f248, %p118;
|
|
mov.b32 %r287, %f851;
|
|
and.b32 %r288, %r287, 8388607;
|
|
or.b32 %r289, %r288, 1065353216;
|
|
mov.b32 %f852, %r289;
|
|
shr.u32 %r290, %r287, 23;
|
|
cvt.rn.f32.u32 %f853, %r290;
|
|
add.f32 %f854, %f850, %f853;
|
|
setp.gt.f32 %p119, %f852, 0f3FB504F3;
|
|
mul.f32 %f855, %f852, 0f3F000000;
|
|
add.f32 %f856, %f854, 0f3F800000;
|
|
selp.f32 %f857, %f855, %f852, %p119;
|
|
selp.f32 %f858, %f856, %f854, %p119;
|
|
add.f32 %f859, %f857, 0fBF800000;
|
|
add.f32 %f848, %f857, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f847,%f848;
|
|
// inline asm
|
|
add.f32 %f860, %f859, %f859;
|
|
mul.f32 %f861, %f847, %f860;
|
|
mul.f32 %f862, %f861, %f861;
|
|
fma.rn.f32 %f865, %f708, %f862, %f707;
|
|
fma.rn.f32 %f867, %f865, %f862, %f710;
|
|
mul.rn.f32 %f868, %f867, %f862;
|
|
mul.rn.f32 %f869, %f868, %f861;
|
|
sub.f32 %f870, %f859, %f861;
|
|
neg.f32 %f871, %f861;
|
|
add.f32 %f872, %f870, %f870;
|
|
fma.rn.f32 %f873, %f871, %f859, %f872;
|
|
mul.rn.f32 %f874, %f847, %f873;
|
|
add.f32 %f875, %f869, %f861;
|
|
sub.f32 %f876, %f861, %f875;
|
|
add.f32 %f877, %f869, %f876;
|
|
add.f32 %f878, %f874, %f877;
|
|
add.f32 %f879, %f875, %f878;
|
|
sub.f32 %f880, %f875, %f879;
|
|
add.f32 %f881, %f878, %f880;
|
|
mul.rn.f32 %f883, %f858, %f726;
|
|
mul.rn.f32 %f885, %f858, %f728;
|
|
add.f32 %f886, %f883, %f879;
|
|
sub.f32 %f887, %f883, %f886;
|
|
add.f32 %f888, %f879, %f887;
|
|
add.f32 %f889, %f881, %f888;
|
|
add.f32 %f890, %f885, %f889;
|
|
add.f32 %f891, %f886, %f890;
|
|
sub.f32 %f892, %f886, %f891;
|
|
add.f32 %f893, %f890, %f892;
|
|
mul.rn.f32 %f895, %f738, %f891;
|
|
neg.f32 %f896, %f895;
|
|
fma.rn.f32 %f897, %f738, %f891, %f896;
|
|
fma.rn.f32 %f898, %f738, %f893, %f897;
|
|
fma.rn.f32 %f900, %f743, %f891, %f898;
|
|
add.rn.f32 %f901, %f895, %f900;
|
|
neg.f32 %f902, %f901;
|
|
add.rn.f32 %f903, %f895, %f902;
|
|
add.rn.f32 %f904, %f903, %f900;
|
|
mov.b32 %r291, %f901;
|
|
setp.eq.s32 %p120, %r291, 1118925336;
|
|
add.s32 %r292, %r291, -1;
|
|
mov.b32 %f905, %r292;
|
|
add.f32 %f906, %f904, 0f37000000;
|
|
selp.f32 %f907, %f905, %f901, %p120;
|
|
selp.f32 %f249, %f906, %f904, %p120;
|
|
mul.f32 %f908, %f907, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f909, %f908;
|
|
fma.rn.f32 %f911, %f909, %f754, %f907;
|
|
fma.rn.f32 %f913, %f909, %f756, %f911;
|
|
mul.f32 %f914, %f913, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f915, %f914;
|
|
add.f32 %f916, %f909, 0f00000000;
|
|
ex2.approx.f32 %f917, %f916;
|
|
mul.f32 %f918, %f915, %f917;
|
|
setp.lt.f32 %p121, %f907, 0fC2D20000;
|
|
selp.f32 %f919, 0f00000000, %f918, %p121;
|
|
setp.gt.f32 %p122, %f907, 0f42D20000;
|
|
selp.f32 %f1272, 0f7F800000, %f919, %p122;
|
|
setp.eq.f32 %p123, %f1272, 0f7F800000;
|
|
@%p123 bra BB0_114;
|
|
|
|
fma.rn.f32 %f1272, %f1272, %f249, %f1272;
|
|
|
|
BB0_114:
|
|
setp.lt.f32 %p124, %f220, 0f00000000;
|
|
and.pred %p6, %p124, %p93;
|
|
mov.b32 %r293, %f1272;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f920, %r294;
|
|
selp.f32 %f1274, %f920, %f1272, %p6;
|
|
setp.eq.f32 %p126, %f220, 0f00000000;
|
|
@%p126 bra BB0_117;
|
|
bra.uni BB0_115;
|
|
|
|
BB0_117:
|
|
add.f32 %f923, %f220, %f220;
|
|
selp.f32 %f1274, %f923, 0f00000000, %p93;
|
|
bra.uni BB0_118;
|
|
|
|
BB0_115:
|
|
setp.geu.f32 %p127, %f220, 0f00000000;
|
|
@%p127 bra BB0_118;
|
|
|
|
cvt.rzi.f32.f32 %f922, %f738;
|
|
setp.neu.f32 %p128, %f922, 0f3EE66666;
|
|
selp.f32 %f1274, 0f7FFFFFFF, %f1274, %p128;
|
|
|
|
BB0_118:
|
|
add.f32 %f924, %f248, 0f3EE66666;
|
|
mov.b32 %r295, %f924;
|
|
setp.lt.s32 %p130, %r295, 2139095040;
|
|
@%p130 bra BB0_123;
|
|
|
|
setp.gtu.f32 %p131, %f248, 0f7F800000;
|
|
@%p131 bra BB0_122;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_122:
|
|
add.f32 %f1274, %f220, 0f3EE66666;
|
|
bra.uni BB0_123;
|
|
|
|
BB0_120:
|
|
setp.neu.f32 %p132, %f248, 0f7F800000;
|
|
@%p132 bra BB0_123;
|
|
|
|
selp.f32 %f1274, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_123:
|
|
setp.eq.f32 %p133, %f220, 0f3F800000;
|
|
selp.f32 %f925, 0f3F800000, %f1274, %p133;
|
|
cvt.u64.u32 %rd93, %r7;
|
|
cvt.u64.u32 %rd92, %r6;
|
|
mov.u64 %rd96, image;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r101, %r102, %rd92, %rd93, %rd27, %rd27);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f926, %f925;
|
|
mul.f32 %f927, %f926, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r298, %f927;
|
|
cvt.sat.f32.f32 %f928, %f247;
|
|
mul.f32 %f929, %f928, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r299, %f929;
|
|
cvt.sat.f32.f32 %f930, %f234;
|
|
mul.f32 %f931, %f930, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r300, %f931;
|
|
cvt.u16.u32 %rs22, %r298;
|
|
cvt.u16.u32 %rs23, %r300;
|
|
cvt.u16.u32 %rs24, %r299;
|
|
mov.u16 %rs25, 255;
|
|
st.v4.u8 [%rd90], {%rs22, %rs24, %rs23, %rs25};
|
|
ld.global.u32 %r480, [imageEnabled];
|
|
|
|
BB0_124:
|
|
cvt.u64.u32 %rd18, %r6;
|
|
cvt.u64.u32 %rd19, %r7;
|
|
and.b32 %r301, %r480, 4;
|
|
setp.eq.s32 %p134, %r301, 0;
|
|
@%p134 bra BB0_128;
|
|
|
|
ld.global.u32 %r302, [additive];
|
|
setp.eq.s32 %p135, %r302, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f490;}
|
|
|
|
// inline asm
|
|
@%p135 bra BB0_127;
|
|
|
|
mov.u64 %rd109, image_HDR;
|
|
cvta.global.u64 %rd98, %rd109;
|
|
mov.u32 %r306, 8;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r101, %r306, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs33, %rs34, %rs35, %rs36}, [%rd97];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f933, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f934, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f935, %rs35;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd98, %r101, %r306, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f936, %f218, %f933;
|
|
add.f32 %f937, %f219, %f934;
|
|
add.f32 %f938, %f220, %f935;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f938;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f937;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f936;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs30, %rs31, %rs32, %rs26};
|
|
bra.uni BB0_128;
|
|
|
|
BB0_127:
|
|
mov.u64 %rd116, image_HDR;
|
|
cvta.global.u64 %rd111, %rd116;
|
|
mov.u32 %r308, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r101, %r308, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f220;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f219;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f218;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs37, %rs38, %rs39, %rs26};
|
|
|
|
BB0_128:
|
|
selp.f32 %f943, 0f40800000, 0f3F800000, %p158;
|
|
mul.f32 %f944, %f943, %f1253;
|
|
mul.f32 %f260, %f943, %f1254;
|
|
mul.f32 %f261, %f943, %f1255;
|
|
mul.f32 %f262, %f943, %f1256;
|
|
mul.f32 %f263, %f943, %f1257;
|
|
mul.f32 %f264, %f943, %f1258;
|
|
mul.f32 %f265, %f943, %f1259;
|
|
mul.f32 %f266, %f943, %f1260;
|
|
mul.f32 %f267, %f943, %f1261;
|
|
ld.global.f32 %f945, [lightColor];
|
|
mul.f32 %f946, %f944, %f945;
|
|
ld.global.f32 %f947, [lightColor+4];
|
|
mul.f32 %f948, %f944, %f947;
|
|
ld.global.f32 %f949, [lightColor+8];
|
|
mul.f32 %f950, %f944, %f949;
|
|
mul.f32 %f268, %f946, 0f3F558750;
|
|
mul.f32 %f269, %f948, 0f3F558750;
|
|
mul.f32 %f270, %f950, 0f3F558750;
|
|
ld.global.u32 %r309, [additive];
|
|
setp.eq.s32 %p136, %r309, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f490;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_130;
|
|
|
|
mov.u64 %rd129, image_RNM0;
|
|
cvta.global.u64 %rd118, %rd129;
|
|
mov.u32 %r313, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r101, %r313, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs47, %rs48, %rs49, %rs50}, [%rd117];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f951, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f952, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f953, %rs49;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd118, %r101, %r313, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f954, %f268, %f951;
|
|
add.f32 %f955, %f269, %f952;
|
|
add.f32 %f956, %f270, %f953;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f956;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f955;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f954;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs44, %rs45, %rs46, %rs40};
|
|
bra.uni BB0_131;
|
|
|
|
BB0_130:
|
|
mov.u64 %rd136, image_RNM0;
|
|
cvta.global.u64 %rd131, %rd136;
|
|
mov.u32 %r315, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r101, %r315, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f270;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f269;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f268;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs51, %rs52, %rs53, %rs40};
|
|
|
|
BB0_131:
|
|
ld.global.f32 %f961, [lightColor];
|
|
mul.f32 %f962, %f260, %f961;
|
|
ld.global.f32 %f963, [lightColor+4];
|
|
mul.f32 %f964, %f260, %f963;
|
|
ld.global.f32 %f965, [lightColor+8];
|
|
mul.f32 %f966, %f260, %f965;
|
|
mul.f32 %f271, %f962, 0f3FB8EBD1;
|
|
mul.f32 %f272, %f964, 0f3FB8EBD1;
|
|
mul.f32 %f273, %f966, 0f3FB8EBD1;
|
|
ld.global.u32 %r316, [additive];
|
|
setp.eq.s32 %p137, %r316, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f490;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_133;
|
|
|
|
mov.u64 %rd149, image_RNM1;
|
|
cvta.global.u64 %rd138, %rd149;
|
|
mov.u32 %r320, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r101, %r320, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs61, %rs62, %rs63, %rs64}, [%rd137];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f967, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f968, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f969, %rs63;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd138, %r101, %r320, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f970, %f271, %f967;
|
|
add.f32 %f971, %f272, %f968;
|
|
add.f32 %f972, %f273, %f969;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f972;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f971;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f970;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs58, %rs59, %rs60, %rs54};
|
|
bra.uni BB0_134;
|
|
|
|
BB0_133:
|
|
mov.u64 %rd156, image_RNM1;
|
|
cvta.global.u64 %rd151, %rd156;
|
|
mov.u32 %r322, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r101, %r322, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f273;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f272;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f271;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs65, %rs66, %rs67, %rs54};
|
|
|
|
BB0_134:
|
|
ld.global.f32 %f977, [lightColor];
|
|
mul.f32 %f978, %f261, %f977;
|
|
ld.global.f32 %f979, [lightColor+4];
|
|
mul.f32 %f980, %f261, %f979;
|
|
ld.global.f32 %f981, [lightColor+8];
|
|
mul.f32 %f982, %f261, %f981;
|
|
mul.f32 %f274, %f978, 0f3FB8EBD1;
|
|
mul.f32 %f275, %f980, 0f3FB8EBD1;
|
|
mul.f32 %f276, %f982, 0f3FB8EBD1;
|
|
ld.global.u32 %r323, [additive];
|
|
setp.eq.s32 %p138, %r323, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f490;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_136;
|
|
|
|
mov.u64 %rd169, image_RNM2;
|
|
cvta.global.u64 %rd158, %rd169;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd158, %r101, %r327, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs75, %rs76, %rs77, %rs78}, [%rd157];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f983, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f984, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f985, %rs77;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd163), _rt_buffer_get_64, (%rd158, %r101, %r327, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f986, %f274, %f983;
|
|
add.f32 %f987, %f275, %f984;
|
|
add.f32 %f988, %f276, %f985;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f988;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f987;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f986;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd163], {%rs72, %rs73, %rs74, %rs68};
|
|
bra.uni BB0_137;
|
|
|
|
BB0_136:
|
|
mov.u64 %rd176, image_RNM2;
|
|
cvta.global.u64 %rd171, %rd176;
|
|
mov.u32 %r329, 8;
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd171, %r101, %r329, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f276;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f275;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f274;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd170], {%rs79, %rs80, %rs81, %rs68};
|
|
|
|
BB0_137:
|
|
ld.global.f32 %f993, [lightColor];
|
|
mul.f32 %f994, %f262, %f993;
|
|
ld.global.f32 %f995, [lightColor+4];
|
|
mul.f32 %f996, %f262, %f995;
|
|
ld.global.f32 %f997, [lightColor+8];
|
|
mul.f32 %f998, %f262, %f997;
|
|
mul.f32 %f277, %f994, 0f3FB8EBD1;
|
|
mul.f32 %f278, %f996, 0f3FB8EBD1;
|
|
mul.f32 %f279, %f998, 0f3FB8EBD1;
|
|
ld.global.u32 %r330, [additive];
|
|
setp.eq.s32 %p139, %r330, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f490;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_139;
|
|
|
|
mov.u64 %rd189, image_RNM3;
|
|
cvta.global.u64 %rd178, %rd189;
|
|
mov.u32 %r334, 8;
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r101, %r334, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd177];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f999, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1000, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1001, %rs91;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd183), _rt_buffer_get_64, (%rd178, %r101, %r334, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1002, %f277, %f999;
|
|
add.f32 %f1003, %f278, %f1000;
|
|
add.f32 %f1004, %f279, %f1001;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f1004;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f1003;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f1002;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd183], {%rs86, %rs87, %rs88, %rs82};
|
|
bra.uni BB0_140;
|
|
|
|
BB0_139:
|
|
mov.u64 %rd196, image_RNM3;
|
|
cvta.global.u64 %rd191, %rd196;
|
|
mov.u32 %r336, 8;
|
|
// inline asm
|
|
call (%rd190), _rt_buffer_get_64, (%rd191, %r101, %r336, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f277;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd190], {%rs93, %rs94, %rs95, %rs82};
|
|
|
|
BB0_140:
|
|
ld.global.f32 %f1009, [lightColor];
|
|
mul.f32 %f1010, %f263, %f1009;
|
|
ld.global.f32 %f1011, [lightColor+4];
|
|
mul.f32 %f1012, %f263, %f1011;
|
|
ld.global.f32 %f1013, [lightColor+8];
|
|
mul.f32 %f1014, %f263, %f1013;
|
|
mul.f32 %f280, %f1010, 0f404EBF87;
|
|
mul.f32 %f281, %f1012, 0f404EBF87;
|
|
mul.f32 %f282, %f1014, 0f404EBF87;
|
|
ld.global.u32 %r337, [additive];
|
|
setp.eq.s32 %p140, %r337, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f490;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_142;
|
|
|
|
mov.u64 %rd209, image_RNM4;
|
|
cvta.global.u64 %rd198, %rd209;
|
|
mov.u32 %r341, 8;
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd198, %r101, %r341, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs103, %rs104, %rs105, %rs106}, [%rd197];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1015, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1016, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1017, %rs105;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd203), _rt_buffer_get_64, (%rd198, %r101, %r341, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1018, %f280, %f1015;
|
|
add.f32 %f1019, %f281, %f1016;
|
|
add.f32 %f1020, %f282, %f1017;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f1020;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1019;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1018;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd203], {%rs100, %rs101, %rs102, %rs96};
|
|
bra.uni BB0_143;
|
|
|
|
BB0_142:
|
|
mov.u64 %rd216, image_RNM4;
|
|
cvta.global.u64 %rd211, %rd216;
|
|
mov.u32 %r343, 8;
|
|
// inline asm
|
|
call (%rd210), _rt_buffer_get_64, (%rd211, %r101, %r343, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f282;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f281;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f280;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd210], {%rs107, %rs108, %rs109, %rs96};
|
|
|
|
BB0_143:
|
|
ld.global.f32 %f1025, [lightColor];
|
|
mul.f32 %f1026, %f264, %f1025;
|
|
ld.global.f32 %f1027, [lightColor+4];
|
|
mul.f32 %f1028, %f264, %f1027;
|
|
ld.global.f32 %f1029, [lightColor+8];
|
|
mul.f32 %f1030, %f264, %f1029;
|
|
mul.f32 %f283, %f1026, 0f404EBF87;
|
|
mul.f32 %f284, %f1028, 0f404EBF87;
|
|
mul.f32 %f285, %f1030, 0f404EBF87;
|
|
ld.global.u32 %r344, [additive];
|
|
setp.eq.s32 %p141, %r344, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f490;}
|
|
|
|
// inline asm
|
|
@%p141 bra BB0_145;
|
|
|
|
mov.u64 %rd229, image_RNM5;
|
|
cvta.global.u64 %rd218, %rd229;
|
|
mov.u32 %r348, 8;
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd218, %r101, %r348, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs117, %rs118, %rs119, %rs120}, [%rd217];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1031, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1032, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1033, %rs119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd223), _rt_buffer_get_64, (%rd218, %r101, %r348, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1034, %f283, %f1031;
|
|
add.f32 %f1035, %f284, %f1032;
|
|
add.f32 %f1036, %f285, %f1033;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f1036;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f1035;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1034;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd223], {%rs114, %rs115, %rs116, %rs110};
|
|
bra.uni BB0_146;
|
|
|
|
BB0_145:
|
|
mov.u64 %rd236, image_RNM5;
|
|
cvta.global.u64 %rd231, %rd236;
|
|
mov.u32 %r350, 8;
|
|
// inline asm
|
|
call (%rd230), _rt_buffer_get_64, (%rd231, %r101, %r350, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f285;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f284;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f283;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd230], {%rs121, %rs122, %rs123, %rs110};
|
|
|
|
BB0_146:
|
|
ld.global.f32 %f1041, [lightColor];
|
|
mul.f32 %f1042, %f265, %f1041;
|
|
ld.global.f32 %f1043, [lightColor+4];
|
|
mul.f32 %f1044, %f265, %f1043;
|
|
ld.global.f32 %f1045, [lightColor+8];
|
|
mul.f32 %f1046, %f265, %f1045;
|
|
mul.f32 %f286, %f1042, 0f40330C94;
|
|
mul.f32 %f287, %f1044, 0f40330C94;
|
|
mul.f32 %f288, %f1046, 0f40330C94;
|
|
ld.global.u32 %r351, [additive];
|
|
setp.eq.s32 %p142, %r351, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f490;}
|
|
|
|
// inline asm
|
|
@%p142 bra BB0_148;
|
|
|
|
mov.u64 %rd249, image_RNM6;
|
|
cvta.global.u64 %rd238, %rd249;
|
|
mov.u32 %r355, 8;
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd238, %r101, %r355, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs131, %rs132, %rs133, %rs134}, [%rd237];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1047, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1048, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1049, %rs133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd243), _rt_buffer_get_64, (%rd238, %r101, %r355, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1050, %f286, %f1047;
|
|
add.f32 %f1051, %f287, %f1048;
|
|
add.f32 %f1052, %f288, %f1049;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f1052;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f1051;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f1050;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd243], {%rs128, %rs129, %rs130, %rs124};
|
|
bra.uni BB0_149;
|
|
|
|
BB0_148:
|
|
mov.u64 %rd256, image_RNM6;
|
|
cvta.global.u64 %rd251, %rd256;
|
|
mov.u32 %r357, 8;
|
|
// inline asm
|
|
call (%rd250), _rt_buffer_get_64, (%rd251, %r101, %r357, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f288;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f287;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f286;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd250], {%rs135, %rs136, %rs137, %rs124};
|
|
|
|
BB0_149:
|
|
ld.global.f32 %f1057, [lightColor];
|
|
mul.f32 %f1058, %f266, %f1057;
|
|
ld.global.f32 %f1059, [lightColor+4];
|
|
mul.f32 %f1060, %f266, %f1059;
|
|
ld.global.f32 %f1061, [lightColor+8];
|
|
mul.f32 %f1062, %f266, %f1061;
|
|
mul.f32 %f289, %f1058, 0f404EBF87;
|
|
mul.f32 %f290, %f1060, 0f404EBF87;
|
|
mul.f32 %f291, %f1062, 0f404EBF87;
|
|
ld.global.u32 %r358, [additive];
|
|
setp.eq.s32 %p143, %r358, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f490;}
|
|
|
|
// inline asm
|
|
@%p143 bra BB0_151;
|
|
|
|
mov.u64 %rd269, image_RNM7;
|
|
cvta.global.u64 %rd258, %rd269;
|
|
mov.u32 %r362, 8;
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd258, %r101, %r362, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd257];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1063, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1064, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1065, %rs147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd263), _rt_buffer_get_64, (%rd258, %r101, %r362, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1066, %f289, %f1063;
|
|
add.f32 %f1067, %f290, %f1064;
|
|
add.f32 %f1068, %f291, %f1065;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1068;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1067;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f1066;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd263], {%rs142, %rs143, %rs144, %rs138};
|
|
bra.uni BB0_152;
|
|
|
|
BB0_151:
|
|
mov.u64 %rd276, image_RNM7;
|
|
cvta.global.u64 %rd271, %rd276;
|
|
mov.u32 %r364, 8;
|
|
// inline asm
|
|
call (%rd270), _rt_buffer_get_64, (%rd271, %r101, %r364, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f290;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f289;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd270], {%rs149, %rs150, %rs151, %rs138};
|
|
|
|
BB0_152:
|
|
ld.global.f32 %f1073, [lightColor];
|
|
mul.f32 %f1074, %f267, %f1073;
|
|
ld.global.f32 %f1075, [lightColor+4];
|
|
mul.f32 %f1076, %f267, %f1075;
|
|
ld.global.f32 %f1077, [lightColor+8];
|
|
mul.f32 %f1078, %f267, %f1077;
|
|
mul.f32 %f292, %f1074, 0f3FCEBF87;
|
|
mul.f32 %f293, %f1076, 0f3FCEBF87;
|
|
mul.f32 %f294, %f1078, 0f3FCEBF87;
|
|
ld.global.u32 %r365, [additive];
|
|
setp.eq.s32 %p144, %r365, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f490;}
|
|
|
|
// inline asm
|
|
@%p144 bra BB0_154;
|
|
|
|
mov.u64 %rd289, image_RNM8;
|
|
cvta.global.u64 %rd278, %rd289;
|
|
mov.u32 %r369, 8;
|
|
// inline asm
|
|
call (%rd277), _rt_buffer_get_64, (%rd278, %r101, %r369, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs159, %rs160, %rs161, %rs162}, [%rd277];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1079, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1080, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1081, %rs161;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd283), _rt_buffer_get_64, (%rd278, %r101, %r369, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1082, %f292, %f1079;
|
|
add.f32 %f1083, %f293, %f1080;
|
|
add.f32 %f1084, %f294, %f1081;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f1084;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1083;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f1082;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd283], {%rs156, %rs157, %rs158, %rs152};
|
|
bra.uni BB0_190;
|
|
|
|
BB0_154:
|
|
mov.u64 %rd296, image_RNM8;
|
|
cvta.global.u64 %rd291, %rd296;
|
|
mov.u32 %r371, 8;
|
|
// inline asm
|
|
call (%rd290), _rt_buffer_get_64, (%rd291, %r101, %r371, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f294;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f292;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd290], {%rs163, %rs164, %rs165, %rs152};
|
|
|
|
BB0_190:
|
|
ret;
|
|
}
|
|
|
|
|