2500 lines
69 KiB
Plaintext
2500 lines
69 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 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 16 .b8 __local_depot0[64];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<150>;
|
|
.reg .b16 %rs<159>;
|
|
.reg .f32 %f<1263>;
|
|
.reg .b32 %r<414>;
|
|
.reg .b64 %rd<320>;
|
|
|
|
|
|
mov.u64 %rd319, __local_depot0;
|
|
cvta.local.u64 %SP, %rd319;
|
|
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 %rd25, %r103;
|
|
cvt.u64.u32 %rd26, %r104;
|
|
mov.u64 %rd29, uvnormal;
|
|
cvta.global.u64 %rd24, %rd29;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd28, 0;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r101, %r102, %rd25, %rd26, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd23];
|
|
shr.u32 %r107, %r5, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r5;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p7, %rs9, 0;
|
|
mov.f32 %f1118, 0f00000000;
|
|
mov.f32 %f1119, %f1118;
|
|
mov.f32 %f1120, %f1118;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd23+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f350, %rs12;
|
|
div.rn.f32 %f351, %f350, 0f437F0000;
|
|
fma.rn.f32 %f352, %f351, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f353, %rs10;
|
|
div.rn.f32 %f354, %f353, 0f437F0000;
|
|
fma.rn.f32 %f355, %f354, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f356, %rs7;
|
|
div.rn.f32 %f357, %f356, 0f437F0000;
|
|
fma.rn.f32 %f358, %f357, 0f40000000, 0fBF800000;
|
|
mul.f32 %f359, %f355, %f355;
|
|
fma.rn.f32 %f360, %f352, %f352, %f359;
|
|
fma.rn.f32 %f361, %f358, %f358, %f360;
|
|
sqrt.rn.f32 %f362, %f361;
|
|
rcp.rn.f32 %f363, %f362;
|
|
mul.f32 %f1118, %f352, %f363;
|
|
mul.f32 %f1119, %f355, %f363;
|
|
mul.f32 %f1120, %f358, %f363;
|
|
|
|
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, %f1119, 0f00000000;
|
|
setp.eq.f32 %p9, %f1118, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1120, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_142;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_142:
|
|
ld.global.u32 %r413, [imageEnabled];
|
|
and.b32 %r337, %r413, 1;
|
|
setp.eq.b32 %p141, %r337, 1;
|
|
@!%p141 bra BB0_144;
|
|
bra.uni BB0_143;
|
|
|
|
BB0_143:
|
|
cvt.u64.u32 %rd200, %r6;
|
|
cvt.u64.u32 %rd201, %r7;
|
|
mov.u64 %rd204, image;
|
|
cvta.global.u64 %rd199, %rd204;
|
|
mov.u64 %rd203, 0;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r101, %r102, %rd200, %rd201, %rd203, %rd203);
|
|
// inline asm
|
|
mov.u16 %rs91, 0;
|
|
st.v4.u8 [%rd198], {%rs91, %rs91, %rs91, %rs91};
|
|
ld.global.u32 %r413, [imageEnabled];
|
|
|
|
BB0_144:
|
|
and.b32 %r340, %r413, 8;
|
|
setp.eq.s32 %p142, %r340, 0;
|
|
@%p142 bra BB0_146;
|
|
|
|
cvt.u64.u32 %rd208, %r7;
|
|
cvt.u64.u32 %rd207, %r6;
|
|
mov.u64 %rd211, image_Mask;
|
|
cvta.global.u64 %rd206, %rd211;
|
|
mov.u64 %rd210, 0;
|
|
// inline asm
|
|
call (%rd205), _rt_buffer_get_64, (%rd206, %r101, %r101, %rd207, %rd208, %rd210, %rd210);
|
|
// inline asm
|
|
mov.f32 %f1075, 0f00000000;
|
|
cvt.rzi.u32.f32 %r343, %f1075;
|
|
cvt.u16.u32 %rs92, %r343;
|
|
mov.u16 %rs93, 0;
|
|
st.v2.u8 [%rd205], {%rs92, %rs93};
|
|
ld.global.u32 %r413, [imageEnabled];
|
|
|
|
BB0_146:
|
|
cvt.u64.u32 %rd21, %r6;
|
|
cvt.u64.u32 %rd22, %r7;
|
|
and.b32 %r344, %r413, 4;
|
|
setp.eq.s32 %p143, %r344, 0;
|
|
@%p143 bra BB0_150;
|
|
|
|
ld.global.u32 %r345, [additive];
|
|
setp.eq.s32 %p144, %r345, 0;
|
|
@%p144 bra BB0_149;
|
|
|
|
mov.u64 %rd224, image_HDR;
|
|
cvta.global.u64 %rd213, %rd224;
|
|
mov.u32 %r349, 8;
|
|
mov.u64 %rd223, 0;
|
|
// inline asm
|
|
call (%rd212), _rt_buffer_get_64, (%rd213, %r101, %r349, %rd21, %rd22, %rd223, %rd223);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd212];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1076, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1077, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1078, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd213, %r101, %r349, %rd21, %rd22, %rd223, %rd223);
|
|
// inline asm
|
|
add.f32 %f1079, %f1076, 0f00000000;
|
|
add.f32 %f1080, %f1077, 0f00000000;
|
|
add.f32 %f1081, %f1078, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1081;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1080;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f1079;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs104, 0;
|
|
st.v4.u16 [%rd218], {%rs97, %rs98, %rs99, %rs104};
|
|
bra.uni BB0_150;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd32, %r121;
|
|
cvt.u64.u32 %rd33, %r122;
|
|
mov.u64 %rd42, uvpos;
|
|
cvta.global.u64 %rd31, %rd42;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r101, %r117, %rd32, %rd33, %rd28, %rd28);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd30+8];
|
|
ld.f32 %f8, [%rd30+4];
|
|
ld.f32 %f7, [%rd30];
|
|
mul.f32 %f380, %f7, 0f3456BF95;
|
|
mul.f32 %f381, %f8, 0f3456BF95;
|
|
mul.f32 %f382, %f9, 0f3456BF95;
|
|
abs.f32 %f383, %f1118;
|
|
div.rn.f32 %f384, %f380, %f383;
|
|
abs.f32 %f385, %f1119;
|
|
div.rn.f32 %f386, %f381, %f385;
|
|
abs.f32 %f387, %f1120;
|
|
div.rn.f32 %f388, %f382, %f387;
|
|
abs.f32 %f389, %f384;
|
|
abs.f32 %f390, %f386;
|
|
abs.f32 %f391, %f388;
|
|
mov.f32 %f392, 0f38D1B717;
|
|
max.f32 %f393, %f389, %f392;
|
|
max.f32 %f394, %f390, %f392;
|
|
max.f32 %f395, %f391, %f392;
|
|
fma.rn.f32 %f10, %f1118, %f393, %f7;
|
|
fma.rn.f32 %f11, %f1119, %f394, %f8;
|
|
fma.rn.f32 %f12, %f1120, %f395, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd38, %r125;
|
|
cvt.u64.u32 %rd39, %r126;
|
|
mov.u64 %rd43, rnd_seeds;
|
|
cvta.global.u64 %rd37, %rd43;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r101, %r102, %rd38, %rd39, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r389, [%rd36];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r385, 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 %r385, 0;
|
|
abs.f32 %f535, %f15;
|
|
abs.f32 %f536, %f14;
|
|
max.f32 %f537, %f536, %f535;
|
|
abs.f32 %f538, %f16;
|
|
max.f32 %f539, %f537, %f538;
|
|
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;
|
|
mov.f32 %f28, %f17;
|
|
mov.f32 %f29, %f17;
|
|
mov.f32 %f30, %f17;
|
|
mov.f32 %f31, %f17;
|
|
mov.f32 %f32, %f17;
|
|
|
|
BB0_6:
|
|
shl.b32 %r10, %r385, 1;
|
|
cvt.s64.s32 %rd46, %r10;
|
|
mov.u64 %rd56, lightMeshBuffer;
|
|
cvta.global.u64 %rd45, %rd56;
|
|
mov.u32 %r130, 1;
|
|
mov.u64 %rd55, 0;
|
|
// inline asm
|
|
call (%rd44), _rt_buffer_get_64, (%rd45, %r130, %r117, %rd46, %rd55, %rd55, %rd55);
|
|
// inline asm
|
|
ld.f32 %f412, [%rd44];
|
|
sub.f32 %f413, %f412, %f7;
|
|
ld.f32 %f414, [%rd44+4];
|
|
sub.f32 %f415, %f414, %f8;
|
|
ld.f32 %f416, [%rd44+8];
|
|
sub.f32 %f417, %f416, %f9;
|
|
mul.f32 %f418, %f415, %f415;
|
|
fma.rn.f32 %f419, %f413, %f413, %f418;
|
|
fma.rn.f32 %f420, %f417, %f417, %f419;
|
|
sqrt.rn.f32 %f33, %f420;
|
|
rcp.rn.f32 %f421, %f33;
|
|
mul.f32 %f34, %f413, %f421;
|
|
mul.f32 %f35, %f415, %f421;
|
|
mul.f32 %f36, %f417, %f421;
|
|
mul.f32 %f422, %f33, %f33;
|
|
div.rn.f32 %f37, %f13, %f422;
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
|
cvt.u64.u32 %rd57, %r135;
|
|
cvt.u64.u32 %rd58, %r134;
|
|
and.b64 %rd52, %rd58, 255;
|
|
and.b64 %rd53, %rd57, 255;
|
|
mov.u64 %rd59, rnd_seeds2;
|
|
cvta.global.u64 %rd51, %rd59;
|
|
mov.u32 %r133, 3;
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd51, %r101, %r133, %rd52, %rd53, %rd55, %rd55);
|
|
// inline asm
|
|
setp.eq.s32 %p15, %r3, 0;
|
|
@%p15 bra BB0_8;
|
|
|
|
ld.u8 %rs14, [%rd50];
|
|
cvt.rn.f32.u16 %f423, %rs14;
|
|
div.rn.f32 %f424, %f423, 0fC37F0000;
|
|
fma.rn.f32 %f425, %f424, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p17, %f37, %f425;
|
|
mov.pred %p149, -1;
|
|
@%p17 bra BB0_23;
|
|
|
|
BB0_8:
|
|
ld.global.f32 %f428, [lightInvCutoff];
|
|
mul.f32 %f38, %f33, %f428;
|
|
mov.f32 %f432, 0f40800000;
|
|
abs.f32 %f40, %f38;
|
|
setp.lt.f32 %p18, %f40, 0f00800000;
|
|
mul.f32 %f434, %f40, 0f4B800000;
|
|
selp.f32 %f435, 0fC3170000, 0fC2FE0000, %p18;
|
|
selp.f32 %f436, %f434, %f40, %p18;
|
|
mov.b32 %r138, %f436;
|
|
and.b32 %r139, %r138, 8388607;
|
|
or.b32 %r140, %r139, 1065353216;
|
|
mov.b32 %f437, %r140;
|
|
shr.u32 %r141, %r138, 23;
|
|
cvt.rn.f32.u32 %f438, %r141;
|
|
add.f32 %f439, %f435, %f438;
|
|
setp.gt.f32 %p19, %f437, 0f3FB504F3;
|
|
mul.f32 %f440, %f437, 0f3F000000;
|
|
add.f32 %f441, %f439, 0f3F800000;
|
|
selp.f32 %f442, %f440, %f437, %p19;
|
|
selp.f32 %f443, %f441, %f439, %p19;
|
|
add.f32 %f444, %f442, 0fBF800000;
|
|
add.f32 %f427, %f442, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f426,%f427;
|
|
// inline asm
|
|
add.f32 %f445, %f444, %f444;
|
|
mul.f32 %f446, %f426, %f445;
|
|
mul.f32 %f447, %f446, %f446;
|
|
mov.f32 %f448, 0f3C4CAF63;
|
|
mov.f32 %f449, 0f3B18F0FE;
|
|
fma.rn.f32 %f450, %f449, %f447, %f448;
|
|
mov.f32 %f451, 0f3DAAAABD;
|
|
fma.rn.f32 %f452, %f450, %f447, %f451;
|
|
mul.rn.f32 %f453, %f452, %f447;
|
|
mul.rn.f32 %f454, %f453, %f446;
|
|
sub.f32 %f455, %f444, %f446;
|
|
neg.f32 %f456, %f446;
|
|
add.f32 %f457, %f455, %f455;
|
|
fma.rn.f32 %f458, %f456, %f444, %f457;
|
|
mul.rn.f32 %f459, %f426, %f458;
|
|
add.f32 %f460, %f454, %f446;
|
|
sub.f32 %f461, %f446, %f460;
|
|
add.f32 %f462, %f454, %f461;
|
|
add.f32 %f463, %f459, %f462;
|
|
add.f32 %f464, %f460, %f463;
|
|
sub.f32 %f465, %f460, %f464;
|
|
add.f32 %f466, %f463, %f465;
|
|
mov.f32 %f467, 0f3F317200;
|
|
mul.rn.f32 %f468, %f443, %f467;
|
|
mov.f32 %f469, 0f35BFBE8E;
|
|
mul.rn.f32 %f470, %f443, %f469;
|
|
add.f32 %f471, %f468, %f464;
|
|
sub.f32 %f472, %f468, %f471;
|
|
add.f32 %f473, %f464, %f472;
|
|
add.f32 %f474, %f466, %f473;
|
|
add.f32 %f475, %f470, %f474;
|
|
add.f32 %f476, %f471, %f475;
|
|
sub.f32 %f477, %f471, %f476;
|
|
add.f32 %f478, %f475, %f477;
|
|
mul.rn.f32 %f479, %f432, %f476;
|
|
neg.f32 %f480, %f479;
|
|
fma.rn.f32 %f481, %f432, %f476, %f480;
|
|
fma.rn.f32 %f482, %f432, %f478, %f481;
|
|
mov.f32 %f483, 0f00000000;
|
|
fma.rn.f32 %f484, %f483, %f476, %f482;
|
|
add.rn.f32 %f485, %f479, %f484;
|
|
neg.f32 %f486, %f485;
|
|
add.rn.f32 %f487, %f479, %f486;
|
|
add.rn.f32 %f488, %f487, %f484;
|
|
mov.b32 %r142, %f485;
|
|
setp.eq.s32 %p20, %r142, 1118925336;
|
|
add.s32 %r143, %r142, -1;
|
|
mov.b32 %f489, %r143;
|
|
add.f32 %f490, %f488, 0f37000000;
|
|
selp.f32 %f491, %f489, %f485, %p20;
|
|
selp.f32 %f41, %f490, %f488, %p20;
|
|
mul.f32 %f492, %f491, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f493, %f492;
|
|
mov.f32 %f494, 0fBF317200;
|
|
fma.rn.f32 %f495, %f493, %f494, %f491;
|
|
mov.f32 %f496, 0fB5BFBE8E;
|
|
fma.rn.f32 %f497, %f493, %f496, %f495;
|
|
mul.f32 %f498, %f497, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f499, %f498;
|
|
add.f32 %f500, %f493, 0f00000000;
|
|
ex2.approx.f32 %f501, %f500;
|
|
mul.f32 %f502, %f499, %f501;
|
|
setp.lt.f32 %p21, %f491, 0fC2D20000;
|
|
selp.f32 %f503, 0f00000000, %f502, %p21;
|
|
setp.gt.f32 %p22, %f491, 0f42D20000;
|
|
selp.f32 %f1137, 0f7F800000, %f503, %p22;
|
|
setp.eq.f32 %p23, %f1137, 0f7F800000;
|
|
@%p23 bra BB0_10;
|
|
|
|
fma.rn.f32 %f1137, %f1137, %f41, %f1137;
|
|
|
|
BB0_10:
|
|
mov.f32 %f1116, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1115, %f1116;
|
|
add.f32 %f1114, %f1115, %f1115;
|
|
mov.f32 %f1113, 0f40800000;
|
|
sub.f32 %f1112, %f1113, %f1114;
|
|
abs.f32 %f1111, %f1112;
|
|
setp.lt.f32 %p24, %f38, 0f00000000;
|
|
setp.eq.f32 %p25, %f1111, 0f3F800000;
|
|
and.pred %p1, %p24, %p25;
|
|
mov.b32 %r144, %f1137;
|
|
xor.b32 %r145, %r144, -2147483648;
|
|
mov.b32 %f504, %r145;
|
|
selp.f32 %f1139, %f504, %f1137, %p1;
|
|
setp.eq.f32 %p26, %f38, 0f00000000;
|
|
@%p26 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f507, %f38, %f38;
|
|
selp.f32 %f1139, %f507, 0f00000000, %p25;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p27, %f38, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
|
|
mov.f32 %f1117, 0f40800000;
|
|
cvt.rzi.f32.f32 %f506, %f1117;
|
|
setp.neu.f32 %p28, %f506, 0f40800000;
|
|
selp.f32 %f1139, 0f7FFFFFFF, %f1139, %p28;
|
|
|
|
BB0_14:
|
|
add.f32 %f508, %f40, 0f40800000;
|
|
mov.b32 %r146, %f508;
|
|
setp.lt.s32 %p30, %r146, 2139095040;
|
|
@%p30 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p31, %f40, 0f7F800000;
|
|
@%p31 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f1139, %f38, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p32, %f40, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
|
|
selp.f32 %f1139, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u64 %rd314, 0;
|
|
mov.u32 %r381, 1;
|
|
mov.u64 %rd313, lightMeshBuffer;
|
|
cvta.global.u64 %rd312, %rd313;
|
|
shl.b32 %r380, %r385, 1;
|
|
mov.f32 %f509, 0f3F800000;
|
|
sub.f32 %f510, %f509, %f1139;
|
|
setp.eq.f32 %p33, %f38, 0f3F800000;
|
|
selp.f32 %f511, 0f00000000, %f510, %p33;
|
|
cvt.sat.f32.f32 %f512, %f511;
|
|
mul.f32 %f513, %f37, %f512;
|
|
add.s32 %r149, %r380, 1;
|
|
cvt.s64.s32 %rd62, %r149;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd312, %r381, %r117, %rd62, %rd314, %rd314, %rd314);
|
|
// inline asm
|
|
ld.f32 %f514, [%rd60];
|
|
mul.f32 %f515, %f34, %f514;
|
|
ld.f32 %f516, [%rd60+4];
|
|
mul.f32 %f517, %f35, %f516;
|
|
neg.f32 %f518, %f517;
|
|
sub.f32 %f519, %f518, %f515;
|
|
ld.f32 %f520, [%rd60+8];
|
|
mul.f32 %f521, %f36, %f520;
|
|
sub.f32 %f522, %f519, %f521;
|
|
cvt.sat.f32.f32 %f523, %f522;
|
|
mul.f32 %f52, %f513, %f523;
|
|
mul.f32 %f524, %f1119, %f35;
|
|
fma.rn.f32 %f525, %f1118, %f34, %f524;
|
|
fma.rn.f32 %f53, %f1120, %f36, %f525;
|
|
setp.leu.f32 %p34, %f52, 0f3727C5AC;
|
|
@%p34 bra BB0_21;
|
|
|
|
mov.u32 %r383, 1;
|
|
cvt.sat.f32.f32 %f534, %f53;
|
|
add.u64 %rd67, %SP, 16;
|
|
cvta.to.local.u64 %rd68, %rd67;
|
|
max.f32 %f532, %f539, %f392;
|
|
sub.f32 %f533, %f33, %f532;
|
|
mov.u32 %r153, 1065353216;
|
|
st.local.u32 [%rd68+8], %r153;
|
|
st.local.u32 [%rd68+4], %r153;
|
|
st.local.u32 [%rd68], %r153;
|
|
ld.global.u32 %r150, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r150, %f10, %f11, %f12, %f34, %f35, %f36, %r383, %f532, %f533, %rd67, %r117);
|
|
// inline asm
|
|
ld.local.f32 %f541, [%rd68];
|
|
mul.f32 %f542, %f52, 0f3EA2F983;
|
|
mul.f32 %f543, %f542, %f541;
|
|
ld.local.f32 %f544, [%rd68+4];
|
|
mul.f32 %f545, %f542, %f544;
|
|
ld.local.f32 %f546, [%rd68+8];
|
|
mul.f32 %f547, %f542, %f546;
|
|
fma.rn.f32 %f20, %f534, %f543, %f20;
|
|
fma.rn.f32 %f19, %f534, %f545, %f19;
|
|
fma.rn.f32 %f18, %f534, %f547, %f18;
|
|
fma.rn.f32 %f26, %f34, %f543, %f26;
|
|
fma.rn.f32 %f25, %f34, %f545, %f25;
|
|
fma.rn.f32 %f24, %f34, %f547, %f24;
|
|
fma.rn.f32 %f29, %f35, %f543, %f29;
|
|
fma.rn.f32 %f28, %f35, %f545, %f28;
|
|
fma.rn.f32 %f27, %f35, %f547, %f27;
|
|
fma.rn.f32 %f32, %f36, %f543, %f32;
|
|
fma.rn.f32 %f31, %f36, %f545, %f31;
|
|
fma.rn.f32 %f30, %f36, %f547, %f30;
|
|
add.f32 %f23, %f23, %f543;
|
|
add.f32 %f22, %f22, %f545;
|
|
add.f32 %f21, %f21, %f547;
|
|
add.f32 %f17, %f17, %f541;
|
|
|
|
BB0_21:
|
|
add.s32 %r385, %r385, 1;
|
|
setp.lt.s32 %p36, %r385, %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;
|
|
mov.f32 %f28, %f17;
|
|
mov.f32 %f29, %f17;
|
|
mov.f32 %f30, %f17;
|
|
mov.f32 %f31, %f17;
|
|
mov.f32 %f32, %f17;
|
|
|
|
BB0_22:
|
|
mov.pred %p149, %p13;
|
|
|
|
BB0_23:
|
|
cvt.rn.f32.s32 %f548, %r385;
|
|
mov.f32 %f549, 0f3F800000;
|
|
max.f32 %f550, %f548, %f549;
|
|
rcp.rn.f32 %f551, %f550;
|
|
mul.f32 %f1247, %f20, %f551;
|
|
mul.f32 %f1248, %f19, %f551;
|
|
mul.f32 %f1249, %f18, %f551;
|
|
div.rn.f32 %f1250, %f17, %f550;
|
|
mul.f32 %f1244, %f23, %f551;
|
|
mul.f32 %f1245, %f22, %f551;
|
|
mul.f32 %f1246, %f21, %f551;
|
|
mul.f32 %f1241, %f26, %f551;
|
|
mul.f32 %f1242, %f25, %f551;
|
|
mul.f32 %f1243, %f24, %f551;
|
|
mul.f32 %f1238, %f29, %f551;
|
|
mul.f32 %f1239, %f28, %f551;
|
|
mul.f32 %f1240, %f27, %f551;
|
|
mul.f32 %f1235, %f32, %f551;
|
|
mul.f32 %f1236, %f31, %f551;
|
|
mul.f32 %f1237, %f30, %f551;
|
|
@!%p149 bra BB0_78;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_24:
|
|
mov.f32 %f1188, 0f00000000;
|
|
setp.lt.s32 %p37, %r3, 1;
|
|
mov.f32 %f1189, %f1188;
|
|
mov.f32 %f1190, %f1188;
|
|
mov.f32 %f1191, %f1188;
|
|
mov.f32 %f1192, %f1188;
|
|
mov.f32 %f1193, %f1188;
|
|
mov.f32 %f1194, %f1188;
|
|
mov.f32 %f1195, %f1188;
|
|
mov.f32 %f1196, %f1188;
|
|
mov.f32 %f1197, %f1188;
|
|
mov.f32 %f1198, %f1188;
|
|
mov.f32 %f1199, %f1188;
|
|
mov.f32 %f1200, %f1188;
|
|
mov.f32 %f1201, %f1188;
|
|
mov.f32 %f1202, %f1188;
|
|
mov.f32 %f1203, %f1188;
|
|
@%p37 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f584, %r4;
|
|
rcp.rn.f32 %f118, %f584;
|
|
add.u64 %rd69, %SP, 28;
|
|
cvta.to.local.u64 %rd3, %rd69;
|
|
mul.f32 %f119, %f10, 0f3456BF95;
|
|
mul.f32 %f120, %f11, 0f3456BF95;
|
|
mul.f32 %f121, %f12, 0f3456BF95;
|
|
add.u64 %rd70, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd70;
|
|
mov.f32 %f1188, 0f00000000;
|
|
mov.u32 %r154, 0;
|
|
abs.f32 %f585, %f120;
|
|
abs.f32 %f586, %f119;
|
|
max.f32 %f587, %f586, %f585;
|
|
abs.f32 %f588, %f121;
|
|
max.f32 %f589, %f587, %f588;
|
|
mov.u32 %r386, %r154;
|
|
mov.f32 %f1189, %f1188;
|
|
mov.f32 %f1190, %f1188;
|
|
mov.f32 %f1191, %f1188;
|
|
mov.f32 %f1192, %f1188;
|
|
mov.f32 %f1193, %f1188;
|
|
mov.f32 %f1194, %f1188;
|
|
mov.f32 %f1195, %f1188;
|
|
mov.f32 %f1196, %f1188;
|
|
mov.f32 %f1197, %f1188;
|
|
mov.f32 %f1198, %f1188;
|
|
mov.f32 %f1199, %f1188;
|
|
mov.f32 %f1200, %f1188;
|
|
mov.f32 %f1201, %f1188;
|
|
mov.f32 %f1202, %f1188;
|
|
mov.f32 %f1203, %f1188;
|
|
|
|
BB0_26:
|
|
cvt.rn.f32.s32 %f138, %r386;
|
|
max.f32 %f139, %f589, %f392;
|
|
mov.u32 %r388, %r154;
|
|
|
|
BB0_27:
|
|
mad.lo.s32 %r156, %r389, 1664525, 1013904223;
|
|
and.b32 %r157, %r156, 16777215;
|
|
cvt.rn.f32.u32 %f591, %r157;
|
|
fma.rn.f32 %f592, %f591, 0f33800000, %f138;
|
|
mul.f32 %f593, %f118, %f592;
|
|
mad.lo.s32 %r389, %r156, 1664525, 1013904223;
|
|
and.b32 %r158, %r389, 16777215;
|
|
cvt.rn.f32.u32 %f594, %r158;
|
|
cvt.rn.f32.s32 %f595, %r388;
|
|
fma.rn.f32 %f596, %f594, 0f33800000, %f595;
|
|
mul.f32 %f597, %f118, %f596;
|
|
fma.rn.f32 %f156, %f593, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f598, %f156, %f156;
|
|
sub.f32 %f600, %f549, %f598;
|
|
mov.f32 %f1216, 0f00000000;
|
|
max.f32 %f602, %f1216, %f600;
|
|
sqrt.rn.f32 %f157, %f602;
|
|
mul.f32 %f1210, %f597, 0f40C90FDB;
|
|
abs.f32 %f159, %f1210;
|
|
setp.neu.f32 %p38, %f159, 0f7F800000;
|
|
mov.f32 %f1204, %f1210;
|
|
@%p38 bra BB0_29;
|
|
|
|
mov.f32 %f603, 0f00000000;
|
|
mul.rn.f32 %f1204, %f1210, %f603;
|
|
|
|
BB0_29:
|
|
mul.f32 %f604, %f1204, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r399, %f604;
|
|
cvt.rn.f32.s32 %f605, %r399;
|
|
neg.f32 %f606, %f605;
|
|
mov.f32 %f607, 0f3FC90FDA;
|
|
fma.rn.f32 %f608, %f606, %f607, %f1204;
|
|
mov.f32 %f609, 0f33A22168;
|
|
fma.rn.f32 %f610, %f606, %f609, %f608;
|
|
mov.f32 %f611, 0f27C234C5;
|
|
fma.rn.f32 %f1205, %f606, %f611, %f610;
|
|
abs.f32 %f612, %f1204;
|
|
setp.leu.f32 %p39, %f612, 0f47CE4780;
|
|
@%p39 bra BB0_40;
|
|
|
|
mov.b32 %r19, %f1204;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r161, %r19, 8;
|
|
or.b32 %r21, %r161, -2147483648;
|
|
mov.u32 %r390, 0;
|
|
mov.u64 %rd316, 0;
|
|
mov.u64 %rd315, %rd3;
|
|
mov.u32 %r391, %r390;
|
|
|
|
BB0_31:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd72, %rd316, 2;
|
|
mov.u64 %rd73, __cudart_i2opi_f;
|
|
add.s64 %rd74, %rd73, %rd72;
|
|
ld.const.u32 %r164, [%rd74];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r162, %r164, %r21, %r391;
|
|
madc.hi.u32 %r391, %r164, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd315], %r162;
|
|
add.s32 %r390, %r390, 1;
|
|
cvt.s64.s32 %rd316, %r390;
|
|
mul.wide.s32 %rd75, %r390, 4;
|
|
add.s64 %rd315, %rd3, %rd75;
|
|
setp.ne.s32 %p40, %r390, 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], %r391;
|
|
mov.u32 %r170, 6;
|
|
sub.s32 %r171, %r170, %r169;
|
|
mul.wide.s32 %rd76, %r171, 4;
|
|
add.s64 %rd10, %rd3, %rd76;
|
|
ld.local.u32 %r392, [%rd10];
|
|
ld.local.u32 %r393, [%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, %r393, %r173;
|
|
shl.b32 %r175, %r392, %r29;
|
|
add.s32 %r392, %r174, %r175;
|
|
ld.local.u32 %r176, [%rd10+-8];
|
|
shr.u32 %r177, %r176, %r173;
|
|
shl.b32 %r178, %r393, %r29;
|
|
add.s32 %r393, %r177, %r178;
|
|
|
|
BB0_34:
|
|
shr.u32 %r179, %r393, 30;
|
|
shl.b32 %r180, %r392, 2;
|
|
add.s32 %r394, %r179, %r180;
|
|
shl.b32 %r35, %r393, 2;
|
|
shr.u32 %r181, %r394, 31;
|
|
shr.u32 %r182, %r392, 30;
|
|
add.s32 %r36, %r181, %r182;
|
|
setp.eq.s32 %p42, %r181, 0;
|
|
@%p42 bra BB0_35;
|
|
bra.uni BB0_36;
|
|
|
|
BB0_35:
|
|
mov.u32 %r395, %r26;
|
|
mov.u32 %r396, %r35;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
not.b32 %r183, %r394;
|
|
neg.s32 %r396, %r35;
|
|
setp.eq.s32 %p43, %r35, 0;
|
|
selp.u32 %r184, 1, 0, %p43;
|
|
add.s32 %r394, %r184, %r183;
|
|
xor.b32 %r395, %r26, -2147483648;
|
|
|
|
BB0_37:
|
|
clz.b32 %r398, %r394;
|
|
setp.eq.s32 %p44, %r398, 0;
|
|
shl.b32 %r185, %r394, %r398;
|
|
mov.u32 %r186, 32;
|
|
sub.s32 %r187, %r186, %r398;
|
|
shr.u32 %r188, %r396, %r187;
|
|
add.s32 %r189, %r188, %r185;
|
|
selp.b32 %r44, %r394, %r189, %p44;
|
|
mov.u32 %r190, -921707870;
|
|
mul.hi.u32 %r397, %r44, %r190;
|
|
setp.eq.s32 %p45, %r26, 0;
|
|
neg.s32 %r191, %r36;
|
|
selp.b32 %r399, %r36, %r191, %p45;
|
|
setp.lt.s32 %p46, %r397, 1;
|
|
@%p46 bra BB0_39;
|
|
|
|
mul.lo.s32 %r192, %r44, -921707870;
|
|
shr.u32 %r193, %r192, 31;
|
|
shl.b32 %r194, %r397, 1;
|
|
add.s32 %r397, %r193, %r194;
|
|
add.s32 %r398, %r398, 1;
|
|
|
|
BB0_39:
|
|
mov.u32 %r195, 126;
|
|
sub.s32 %r196, %r195, %r398;
|
|
shl.b32 %r197, %r196, 23;
|
|
add.s32 %r198, %r397, 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, %r395;
|
|
mov.b32 %f1205, %r203;
|
|
|
|
BB0_40:
|
|
mul.rn.f32 %f165, %f1205, %f1205;
|
|
add.s32 %r52, %r399, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p47, %r53, 0;
|
|
@%p47 bra BB0_42;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_42:
|
|
mov.f32 %f615, 0f3C08839E;
|
|
mov.f32 %f616, 0fB94CA1F9;
|
|
fma.rn.f32 %f1206, %f616, %f165, %f615;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_41:
|
|
mov.f32 %f613, 0fBAB6061A;
|
|
mov.f32 %f614, 0f37CCF5CE;
|
|
fma.rn.f32 %f1206, %f614, %f165, %f613;
|
|
|
|
BB0_43:
|
|
@%p47 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f620, 0fBE2AAAA3;
|
|
fma.rn.f32 %f621, %f1206, %f165, %f620;
|
|
mov.f32 %f622, 0f00000000;
|
|
fma.rn.f32 %f1207, %f621, %f165, %f622;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f617, 0f3D2AAAA5;
|
|
fma.rn.f32 %f618, %f1206, %f165, %f617;
|
|
mov.f32 %f619, 0fBF000000;
|
|
fma.rn.f32 %f1207, %f618, %f165, %f619;
|
|
|
|
BB0_46:
|
|
fma.rn.f32 %f1208, %f1207, %f1205, %f1205;
|
|
@%p47 bra BB0_48;
|
|
|
|
fma.rn.f32 %f1208, %f1207, %f165, %f549;
|
|
|
|
BB0_48:
|
|
and.b32 %r204, %r52, 2;
|
|
setp.eq.s32 %p50, %r204, 0;
|
|
@%p50 bra BB0_50;
|
|
|
|
mov.f32 %f625, 0fBF800000;
|
|
fma.rn.f32 %f1208, %f1208, %f625, %f1216;
|
|
|
|
BB0_50:
|
|
mul.f32 %f177, %f157, %f1208;
|
|
@%p38 bra BB0_52;
|
|
|
|
mov.f32 %f626, 0f00000000;
|
|
mul.rn.f32 %f1210, %f1210, %f626;
|
|
|
|
BB0_52:
|
|
mul.f32 %f627, %f1210, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r409, %f627;
|
|
cvt.rn.f32.s32 %f628, %r409;
|
|
neg.f32 %f629, %f628;
|
|
fma.rn.f32 %f631, %f629, %f607, %f1210;
|
|
fma.rn.f32 %f633, %f629, %f609, %f631;
|
|
fma.rn.f32 %f1211, %f629, %f611, %f633;
|
|
abs.f32 %f635, %f1210;
|
|
setp.leu.f32 %p52, %f635, 0f47CE4780;
|
|
@%p52 bra BB0_63;
|
|
|
|
mov.b32 %r55, %f1210;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r207, %r55, 8;
|
|
or.b32 %r57, %r207, -2147483648;
|
|
mov.u32 %r400, 0;
|
|
mov.u64 %rd317, %rd3;
|
|
mov.u64 %rd318, %rd28;
|
|
mov.u32 %r401, %r400;
|
|
|
|
BB0_54:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd318, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r210, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r208, %r210, %r57, %r401;
|
|
madc.hi.u32 %r401, %r210, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd317], %r208;
|
|
add.s32 %r400, %r400, 1;
|
|
cvt.s64.s32 %rd318, %r400;
|
|
mul.wide.s32 %rd81, %r400, 4;
|
|
add.s64 %rd317, %rd3, %rd81;
|
|
setp.ne.s32 %p53, %r400, 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], %r401;
|
|
mov.u32 %r216, 6;
|
|
sub.s32 %r217, %r216, %r215;
|
|
mul.wide.s32 %rd82, %r217, 4;
|
|
add.s64 %rd16, %rd3, %rd82;
|
|
ld.local.u32 %r402, [%rd16];
|
|
ld.local.u32 %r403, [%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, %r403, %r219;
|
|
shl.b32 %r221, %r402, %r65;
|
|
add.s32 %r402, %r220, %r221;
|
|
ld.local.u32 %r222, [%rd16+-8];
|
|
shr.u32 %r223, %r222, %r219;
|
|
shl.b32 %r224, %r403, %r65;
|
|
add.s32 %r403, %r223, %r224;
|
|
|
|
BB0_57:
|
|
shr.u32 %r225, %r403, 30;
|
|
shl.b32 %r226, %r402, 2;
|
|
add.s32 %r404, %r225, %r226;
|
|
shl.b32 %r71, %r403, 2;
|
|
shr.u32 %r227, %r404, 31;
|
|
shr.u32 %r228, %r402, 30;
|
|
add.s32 %r72, %r227, %r228;
|
|
setp.eq.s32 %p55, %r227, 0;
|
|
@%p55 bra BB0_58;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_58:
|
|
mov.u32 %r405, %r62;
|
|
mov.u32 %r406, %r71;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
not.b32 %r229, %r404;
|
|
neg.s32 %r406, %r71;
|
|
setp.eq.s32 %p56, %r71, 0;
|
|
selp.u32 %r230, 1, 0, %p56;
|
|
add.s32 %r404, %r230, %r229;
|
|
xor.b32 %r405, %r62, -2147483648;
|
|
|
|
BB0_60:
|
|
clz.b32 %r408, %r404;
|
|
setp.eq.s32 %p57, %r408, 0;
|
|
shl.b32 %r231, %r404, %r408;
|
|
mov.u32 %r232, 32;
|
|
sub.s32 %r233, %r232, %r408;
|
|
shr.u32 %r234, %r406, %r233;
|
|
add.s32 %r235, %r234, %r231;
|
|
selp.b32 %r80, %r404, %r235, %p57;
|
|
mov.u32 %r236, -921707870;
|
|
mul.hi.u32 %r407, %r80, %r236;
|
|
setp.eq.s32 %p58, %r62, 0;
|
|
neg.s32 %r237, %r72;
|
|
selp.b32 %r409, %r72, %r237, %p58;
|
|
setp.lt.s32 %p59, %r407, 1;
|
|
@%p59 bra BB0_62;
|
|
|
|
mul.lo.s32 %r238, %r80, -921707870;
|
|
shr.u32 %r239, %r238, 31;
|
|
shl.b32 %r240, %r407, 1;
|
|
add.s32 %r407, %r239, %r240;
|
|
add.s32 %r408, %r408, 1;
|
|
|
|
BB0_62:
|
|
mov.u32 %r241, 126;
|
|
sub.s32 %r242, %r241, %r408;
|
|
shl.b32 %r243, %r242, 23;
|
|
add.s32 %r244, %r407, 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, %r405;
|
|
mov.b32 %f1211, %r249;
|
|
|
|
BB0_63:
|
|
mul.rn.f32 %f183, %f1211, %f1211;
|
|
and.b32 %r88, %r409, 1;
|
|
setp.eq.s32 %p60, %r88, 0;
|
|
@%p60 bra BB0_65;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_65:
|
|
mov.f32 %f638, 0f3C08839E;
|
|
mov.f32 %f639, 0fB94CA1F9;
|
|
fma.rn.f32 %f1212, %f639, %f183, %f638;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_64:
|
|
mov.f32 %f636, 0fBAB6061A;
|
|
mov.f32 %f637, 0f37CCF5CE;
|
|
fma.rn.f32 %f1212, %f637, %f183, %f636;
|
|
|
|
BB0_66:
|
|
@%p60 bra BB0_68;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_68:
|
|
mov.f32 %f643, 0fBE2AAAA3;
|
|
fma.rn.f32 %f644, %f1212, %f183, %f643;
|
|
mov.f32 %f645, 0f00000000;
|
|
fma.rn.f32 %f1213, %f644, %f183, %f645;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_67:
|
|
mov.f32 %f640, 0f3D2AAAA5;
|
|
fma.rn.f32 %f641, %f1212, %f183, %f640;
|
|
mov.f32 %f642, 0fBF000000;
|
|
fma.rn.f32 %f1213, %f641, %f183, %f642;
|
|
|
|
BB0_69:
|
|
fma.rn.f32 %f1214, %f1213, %f1211, %f1211;
|
|
@%p60 bra BB0_71;
|
|
|
|
fma.rn.f32 %f1214, %f1213, %f183, %f549;
|
|
|
|
BB0_71:
|
|
and.b32 %r250, %r409, 2;
|
|
setp.eq.s32 %p63, %r250, 0;
|
|
@%p63 bra BB0_73;
|
|
|
|
mov.f32 %f648, 0fBF800000;
|
|
fma.rn.f32 %f1214, %f1214, %f648, %f1216;
|
|
|
|
BB0_73:
|
|
mov.f32 %f660, 0f7F7FFFFF;
|
|
st.local.v4.f32 [%rd4], {%f549, %f549, %f549, %f660};
|
|
ld.global.u32 %r251, [root];
|
|
mul.f32 %f653, %f157, %f1214;
|
|
mov.u32 %r252, 0;
|
|
mov.f32 %f656, 0f6C4ECB8F;
|
|
mov.u32 %r253, 16;
|
|
// inline asm
|
|
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f177, %f653, %f156, %r252, %f139, %f656, %rd70, %r253);
|
|
// inline asm
|
|
ld.local.f32 %f196, [%rd4+12];
|
|
setp.ge.f32 %p64, %f196, 0f00000000;
|
|
mov.f32 %f1217, %f1216;
|
|
mov.f32 %f1218, %f1216;
|
|
@%p64 bra BB0_75;
|
|
|
|
ld.local.v4.f32 {%f1218, %f1217, %f1216, %f665}, [%rd4];
|
|
|
|
BB0_75:
|
|
setp.ltu.f32 %p65, %f196, 0f00000000;
|
|
selp.f32 %f666, 0f3F800000, 0f00000000, %p65;
|
|
add.f32 %f1188, %f1188, %f666;
|
|
fma.rn.f32 %f1197, %f177, %f1218, %f1197;
|
|
fma.rn.f32 %f1196, %f177, %f1217, %f1196;
|
|
fma.rn.f32 %f1195, %f177, %f1216, %f1195;
|
|
fma.rn.f32 %f1200, %f653, %f1218, %f1200;
|
|
fma.rn.f32 %f1199, %f653, %f1217, %f1199;
|
|
fma.rn.f32 %f1198, %f653, %f1216, %f1198;
|
|
fma.rn.f32 %f1203, %f156, %f1218, %f1203;
|
|
fma.rn.f32 %f1202, %f156, %f1217, %f1202;
|
|
fma.rn.f32 %f1201, %f156, %f1216, %f1201;
|
|
add.f32 %f1194, %f1194, %f1218;
|
|
add.f32 %f1193, %f1193, %f1217;
|
|
add.f32 %f1192, %f1192, %f1216;
|
|
mul.f32 %f667, %f1119, %f653;
|
|
fma.rn.f32 %f668, %f1118, %f177, %f667;
|
|
fma.rn.f32 %f669, %f1120, %f156, %f668;
|
|
cvt.sat.f32.f32 %f670, %f669;
|
|
fma.rn.f32 %f1191, %f1218, %f670, %f1191;
|
|
fma.rn.f32 %f1190, %f1217, %f670, %f1190;
|
|
fma.rn.f32 %f1189, %f1216, %f670, %f1189;
|
|
add.s32 %r388, %r388, 1;
|
|
setp.lt.s32 %p66, %r388, %r4;
|
|
@%p66 bra BB0_27;
|
|
|
|
add.s32 %r386, %r386, 1;
|
|
setp.lt.s32 %p67, %r386, %r4;
|
|
@%p67 bra BB0_26;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r254, %r4, %r4;
|
|
cvt.rn.f32.s32 %f671, %r254;
|
|
rcp.rn.f32 %f672, %f671;
|
|
mul.f32 %f673, %f1191, %f672;
|
|
mul.f32 %f674, %f1190, %f672;
|
|
mul.f32 %f675, %f1189, %f672;
|
|
div.rn.f32 %f1250, %f1188, %f671;
|
|
mul.f32 %f1244, %f1194, %f672;
|
|
mul.f32 %f1245, %f1193, %f672;
|
|
mul.f32 %f1246, %f1192, %f672;
|
|
mul.f32 %f1241, %f1197, %f672;
|
|
mul.f32 %f1242, %f1196, %f672;
|
|
mul.f32 %f1243, %f1195, %f672;
|
|
mul.f32 %f1238, %f1200, %f672;
|
|
mul.f32 %f1239, %f1199, %f672;
|
|
mul.f32 %f1240, %f1198, %f672;
|
|
mul.f32 %f1235, %f1203, %f672;
|
|
mul.f32 %f1236, %f1202, %f672;
|
|
mul.f32 %f1237, %f1201, %f672;
|
|
fma.rn.f32 %f1247, %f1191, %f672, %f673;
|
|
fma.rn.f32 %f1248, %f1190, %f672, %f674;
|
|
fma.rn.f32 %f1249, %f1189, %f672, %f675;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p68, %r1, -1;
|
|
selp.f32 %f267, %f1250, %f1247, %p68;
|
|
ld.global.u32 %r411, [imageEnabled];
|
|
and.b32 %r255, %r411, 8;
|
|
setp.eq.s32 %p69, %r255, 0;
|
|
@%p69 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd86, %r6;
|
|
cvt.u64.u32 %rd87, %r7;
|
|
mov.u64 %rd90, image_Mask;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r101, %r101, %rd86, %rd87, %rd28, %rd28);
|
|
// inline asm
|
|
mov.f32 %f678, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f679, %f678;
|
|
fma.rn.f32 %f680, %f679, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f268, %f680;
|
|
abs.f32 %f269, %f267;
|
|
setp.lt.f32 %p70, %f269, 0f00800000;
|
|
mul.f32 %f681, %f269, 0f4B800000;
|
|
selp.f32 %f682, 0fC3170000, 0fC2FE0000, %p70;
|
|
selp.f32 %f683, %f681, %f269, %p70;
|
|
mov.b32 %r258, %f683;
|
|
and.b32 %r259, %r258, 8388607;
|
|
or.b32 %r260, %r259, 1065353216;
|
|
mov.b32 %f684, %r260;
|
|
shr.u32 %r261, %r258, 23;
|
|
cvt.rn.f32.u32 %f685, %r261;
|
|
add.f32 %f686, %f682, %f685;
|
|
setp.gt.f32 %p71, %f684, 0f3FB504F3;
|
|
mul.f32 %f687, %f684, 0f3F000000;
|
|
add.f32 %f688, %f686, 0f3F800000;
|
|
selp.f32 %f689, %f687, %f684, %p71;
|
|
selp.f32 %f690, %f688, %f686, %p71;
|
|
add.f32 %f691, %f689, 0fBF800000;
|
|
add.f32 %f677, %f689, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f676,%f677;
|
|
// inline asm
|
|
add.f32 %f692, %f691, %f691;
|
|
mul.f32 %f693, %f676, %f692;
|
|
mul.f32 %f694, %f693, %f693;
|
|
mov.f32 %f695, 0f3C4CAF63;
|
|
mov.f32 %f696, 0f3B18F0FE;
|
|
fma.rn.f32 %f697, %f696, %f694, %f695;
|
|
mov.f32 %f698, 0f3DAAAABD;
|
|
fma.rn.f32 %f699, %f697, %f694, %f698;
|
|
mul.rn.f32 %f700, %f699, %f694;
|
|
mul.rn.f32 %f701, %f700, %f693;
|
|
sub.f32 %f702, %f691, %f693;
|
|
neg.f32 %f703, %f693;
|
|
add.f32 %f704, %f702, %f702;
|
|
fma.rn.f32 %f705, %f703, %f691, %f704;
|
|
mul.rn.f32 %f706, %f676, %f705;
|
|
add.f32 %f707, %f701, %f693;
|
|
sub.f32 %f708, %f693, %f707;
|
|
add.f32 %f709, %f701, %f708;
|
|
add.f32 %f710, %f706, %f709;
|
|
add.f32 %f711, %f707, %f710;
|
|
sub.f32 %f712, %f707, %f711;
|
|
add.f32 %f713, %f710, %f712;
|
|
mov.f32 %f714, 0f3F317200;
|
|
mul.rn.f32 %f715, %f690, %f714;
|
|
mov.f32 %f716, 0f35BFBE8E;
|
|
mul.rn.f32 %f717, %f690, %f716;
|
|
add.f32 %f718, %f715, %f711;
|
|
sub.f32 %f719, %f715, %f718;
|
|
add.f32 %f720, %f711, %f719;
|
|
add.f32 %f721, %f713, %f720;
|
|
add.f32 %f722, %f717, %f721;
|
|
add.f32 %f723, %f718, %f722;
|
|
sub.f32 %f724, %f718, %f723;
|
|
add.f32 %f725, %f722, %f724;
|
|
mov.f32 %f726, 0f3EE8BA2E;
|
|
mul.rn.f32 %f727, %f726, %f723;
|
|
neg.f32 %f728, %f727;
|
|
fma.rn.f32 %f729, %f726, %f723, %f728;
|
|
fma.rn.f32 %f730, %f726, %f725, %f729;
|
|
mov.f32 %f731, 0f00000000;
|
|
fma.rn.f32 %f732, %f731, %f723, %f730;
|
|
add.rn.f32 %f733, %f727, %f732;
|
|
neg.f32 %f734, %f733;
|
|
add.rn.f32 %f735, %f727, %f734;
|
|
add.rn.f32 %f736, %f735, %f732;
|
|
mov.b32 %r262, %f733;
|
|
setp.eq.s32 %p72, %r262, 1118925336;
|
|
add.s32 %r263, %r262, -1;
|
|
mov.b32 %f737, %r263;
|
|
add.f32 %f738, %f736, 0f37000000;
|
|
selp.f32 %f739, %f737, %f733, %p72;
|
|
selp.f32 %f270, %f738, %f736, %p72;
|
|
mul.f32 %f740, %f739, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f741, %f740;
|
|
mov.f32 %f742, 0fBF317200;
|
|
fma.rn.f32 %f743, %f741, %f742, %f739;
|
|
mov.f32 %f744, 0fB5BFBE8E;
|
|
fma.rn.f32 %f745, %f741, %f744, %f743;
|
|
mul.f32 %f746, %f745, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f747, %f746;
|
|
add.f32 %f748, %f741, 0f00000000;
|
|
ex2.approx.f32 %f749, %f748;
|
|
mul.f32 %f750, %f747, %f749;
|
|
setp.lt.f32 %p73, %f739, 0fC2D20000;
|
|
selp.f32 %f751, 0f00000000, %f750, %p73;
|
|
setp.gt.f32 %p74, %f739, 0f42D20000;
|
|
selp.f32 %f1251, 0f7F800000, %f751, %p74;
|
|
setp.eq.f32 %p75, %f1251, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
|
|
fma.rn.f32 %f1251, %f1251, %f270, %f1251;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p76, %f267, 0f00000000;
|
|
setp.eq.f32 %p77, %f268, 0f3F800000;
|
|
and.pred %p3, %p76, %p77;
|
|
mov.b32 %r264, %f1251;
|
|
xor.b32 %r265, %r264, -2147483648;
|
|
mov.b32 %f752, %r265;
|
|
selp.f32 %f1253, %f752, %f1251, %p3;
|
|
setp.eq.f32 %p78, %f267, 0f00000000;
|
|
@%p78 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f755, %f267, %f267;
|
|
selp.f32 %f1253, %f755, 0f00000000, %p77;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_149:
|
|
mov.u64 %rd231, image_HDR;
|
|
cvta.global.u64 %rd226, %rd231;
|
|
mov.u32 %r351, 8;
|
|
mov.u64 %rd230, 0;
|
|
// inline asm
|
|
call (%rd225), _rt_buffer_get_64, (%rd226, %r101, %r351, %rd21, %rd22, %rd230, %rd230);
|
|
// inline asm
|
|
mov.f32 %f1082, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1082;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd225], {%rs105, %rs105, %rs105, %rs106};
|
|
|
|
BB0_150:
|
|
ld.global.u32 %r352, [additive];
|
|
setp.eq.s32 %p145, %r352, 0;
|
|
@%p145 bra BB0_152;
|
|
|
|
mov.u64 %rd244, image_RNM0;
|
|
cvta.global.u64 %rd233, %rd244;
|
|
mov.u32 %r356, 8;
|
|
mov.u64 %rd243, 0;
|
|
// inline asm
|
|
call (%rd232), _rt_buffer_get_64, (%rd233, %r101, %r356, %rd21, %rd22, %rd243, %rd243);
|
|
// inline asm
|
|
ld.v4.u16 {%rs113, %rs114, %rs115, %rs116}, [%rd232];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1083, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1084, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1085, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd233, %r101, %r356, %rd21, %rd22, %rd243, %rd243);
|
|
// inline asm
|
|
add.f32 %f1086, %f1083, 0f00000000;
|
|
add.f32 %f1087, %f1084, 0f00000000;
|
|
add.f32 %f1088, %f1085, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1088;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1087;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1086;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs117, 0;
|
|
st.v4.u16 [%rd238], {%rs110, %rs111, %rs112, %rs117};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd251, image_RNM0;
|
|
cvta.global.u64 %rd246, %rd251;
|
|
mov.u32 %r358, 8;
|
|
mov.u64 %rd250, 0;
|
|
// inline asm
|
|
call (%rd245), _rt_buffer_get_64, (%rd246, %r101, %r358, %rd21, %rd22, %rd250, %rd250);
|
|
// inline asm
|
|
mov.f32 %f1089, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1089;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd245], {%rs118, %rs118, %rs118, %rs119};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r359, [additive];
|
|
setp.eq.s32 %p146, %r359, 0;
|
|
@%p146 bra BB0_155;
|
|
|
|
mov.u64 %rd264, image_RNM1;
|
|
cvta.global.u64 %rd253, %rd264;
|
|
mov.u32 %r363, 8;
|
|
mov.u64 %rd263, 0;
|
|
// inline asm
|
|
call (%rd252), _rt_buffer_get_64, (%rd253, %r101, %r363, %rd21, %rd22, %rd263, %rd263);
|
|
// inline asm
|
|
ld.v4.u16 {%rs126, %rs127, %rs128, %rs129}, [%rd252];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1090, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1091, %rs127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1092, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd253, %r101, %r363, %rd21, %rd22, %rd263, %rd263);
|
|
// inline asm
|
|
add.f32 %f1093, %f1090, 0f00000000;
|
|
add.f32 %f1094, %f1091, 0f00000000;
|
|
add.f32 %f1095, %f1092, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1095;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1094;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1093;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs130, 0;
|
|
st.v4.u16 [%rd258], {%rs123, %rs124, %rs125, %rs130};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd271, image_RNM1;
|
|
cvta.global.u64 %rd266, %rd271;
|
|
mov.u32 %r365, 8;
|
|
mov.u64 %rd270, 0;
|
|
// inline asm
|
|
call (%rd265), _rt_buffer_get_64, (%rd266, %r101, %r365, %rd21, %rd22, %rd270, %rd270);
|
|
// inline asm
|
|
mov.f32 %f1096, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1096;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd265], {%rs131, %rs131, %rs131, %rs132};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r366, [additive];
|
|
setp.eq.s32 %p147, %r366, 0;
|
|
@%p147 bra BB0_158;
|
|
|
|
mov.u64 %rd284, image_RNM2;
|
|
cvta.global.u64 %rd273, %rd284;
|
|
mov.u32 %r370, 8;
|
|
mov.u64 %rd283, 0;
|
|
// inline asm
|
|
call (%rd272), _rt_buffer_get_64, (%rd273, %r101, %r370, %rd21, %rd22, %rd283, %rd283);
|
|
// inline asm
|
|
ld.v4.u16 {%rs139, %rs140, %rs141, %rs142}, [%rd272];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1097, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1098, %rs140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1099, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd273, %r101, %r370, %rd21, %rd22, %rd283, %rd283);
|
|
// inline asm
|
|
add.f32 %f1100, %f1097, 0f00000000;
|
|
add.f32 %f1101, %f1098, 0f00000000;
|
|
add.f32 %f1102, %f1099, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1100;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs143, 0;
|
|
st.v4.u16 [%rd278], {%rs136, %rs137, %rs138, %rs143};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd291, image_RNM2;
|
|
cvta.global.u64 %rd286, %rd291;
|
|
mov.u32 %r372, 8;
|
|
mov.u64 %rd290, 0;
|
|
// inline asm
|
|
call (%rd285), _rt_buffer_get_64, (%rd286, %r101, %r372, %rd21, %rd22, %rd290, %rd290);
|
|
// inline asm
|
|
mov.f32 %f1103, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1103;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd285], {%rs144, %rs144, %rs144, %rs145};
|
|
|
|
BB0_159:
|
|
ld.global.u32 %r373, [additive];
|
|
setp.eq.s32 %p148, %r373, 0;
|
|
@%p148 bra BB0_161;
|
|
|
|
mov.u64 %rd304, image_RNM3;
|
|
cvta.global.u64 %rd293, %rd304;
|
|
mov.u32 %r377, 8;
|
|
mov.u64 %rd303, 0;
|
|
// inline asm
|
|
call (%rd292), _rt_buffer_get_64, (%rd293, %r101, %r377, %rd21, %rd22, %rd303, %rd303);
|
|
// inline asm
|
|
ld.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd292];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1104, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1105, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1106, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd293, %r101, %r377, %rd21, %rd22, %rd303, %rd303);
|
|
// inline asm
|
|
add.f32 %f1107, %f1104, 0f00000000;
|
|
add.f32 %f1108, %f1105, 0f00000000;
|
|
add.f32 %f1109, %f1106, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1109;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1108;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1107;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs156, 0;
|
|
st.v4.u16 [%rd298], {%rs149, %rs150, %rs151, %rs156};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_161:
|
|
mov.u64 %rd311, image_RNM3;
|
|
cvta.global.u64 %rd306, %rd311;
|
|
mov.u32 %r379, 8;
|
|
mov.u64 %rd310, 0;
|
|
// inline asm
|
|
call (%rd305), _rt_buffer_get_64, (%rd306, %r101, %r379, %rd21, %rd22, %rd310, %rd310);
|
|
// inline asm
|
|
mov.f32 %f1110, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1110;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd305], {%rs157, %rs157, %rs157, %rs158};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p79, %f267, 0f00000000;
|
|
@%p79 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f754, %f726;
|
|
setp.neu.f32 %p80, %f754, 0f3EE8BA2E;
|
|
selp.f32 %f1253, 0f7FFFFFFF, %f1253, %p80;
|
|
|
|
BB0_85:
|
|
add.f32 %f756, %f269, 0f3EE8BA2E;
|
|
mov.b32 %r266, %f756;
|
|
setp.lt.s32 %p82, %r266, 2139095040;
|
|
@%p82 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p83, %f269, 0f7F800000;
|
|
@%p83 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f1253, %f267, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p84, %f269, 0f7F800000;
|
|
@%p84 bra BB0_90;
|
|
|
|
selp.f32 %f1253, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f757, %f1253, 0f437F0000;
|
|
setp.eq.f32 %p85, %f267, 0f3F800000;
|
|
selp.f32 %f758, 0f437F0000, %f757, %p85;
|
|
cvt.rzi.u32.f32 %r267, %f758;
|
|
cvt.u16.u32 %rs15, %r267;
|
|
mov.u16 %rs16, 255;
|
|
st.v2.u8 [%rd84], {%rs15, %rs16};
|
|
ld.global.u32 %r411, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f759, [lightColor];
|
|
mul.f32 %f281, %f1247, %f759;
|
|
ld.global.f32 %f760, [lightColor+4];
|
|
mul.f32 %f282, %f1248, %f760;
|
|
ld.global.f32 %f761, [lightColor+8];
|
|
mul.f32 %f283, %f1249, %f761;
|
|
and.b32 %r268, %r411, 1;
|
|
setp.eq.b32 %p86, %r268, 1;
|
|
@!%p86 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f764, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f765, %f764;
|
|
fma.rn.f32 %f766, %f765, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f284, %f766;
|
|
abs.f32 %f285, %f281;
|
|
setp.lt.f32 %p87, %f285, 0f00800000;
|
|
mul.f32 %f767, %f285, 0f4B800000;
|
|
selp.f32 %f768, 0fC3170000, 0fC2FE0000, %p87;
|
|
selp.f32 %f769, %f767, %f285, %p87;
|
|
mov.b32 %r269, %f769;
|
|
and.b32 %r270, %r269, 8388607;
|
|
or.b32 %r271, %r270, 1065353216;
|
|
mov.b32 %f770, %r271;
|
|
shr.u32 %r272, %r269, 23;
|
|
cvt.rn.f32.u32 %f771, %r272;
|
|
add.f32 %f772, %f768, %f771;
|
|
setp.gt.f32 %p88, %f770, 0f3FB504F3;
|
|
mul.f32 %f773, %f770, 0f3F000000;
|
|
add.f32 %f774, %f772, 0f3F800000;
|
|
selp.f32 %f775, %f773, %f770, %p88;
|
|
selp.f32 %f776, %f774, %f772, %p88;
|
|
add.f32 %f777, %f775, 0fBF800000;
|
|
add.f32 %f763, %f775, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f762,%f763;
|
|
// inline asm
|
|
add.f32 %f778, %f777, %f777;
|
|
mul.f32 %f779, %f762, %f778;
|
|
mul.f32 %f780, %f779, %f779;
|
|
mov.f32 %f781, 0f3C4CAF63;
|
|
mov.f32 %f782, 0f3B18F0FE;
|
|
fma.rn.f32 %f783, %f782, %f780, %f781;
|
|
mov.f32 %f784, 0f3DAAAABD;
|
|
fma.rn.f32 %f785, %f783, %f780, %f784;
|
|
mul.rn.f32 %f786, %f785, %f780;
|
|
mul.rn.f32 %f787, %f786, %f779;
|
|
sub.f32 %f788, %f777, %f779;
|
|
neg.f32 %f789, %f779;
|
|
add.f32 %f790, %f788, %f788;
|
|
fma.rn.f32 %f791, %f789, %f777, %f790;
|
|
mul.rn.f32 %f792, %f762, %f791;
|
|
add.f32 %f793, %f787, %f779;
|
|
sub.f32 %f794, %f779, %f793;
|
|
add.f32 %f795, %f787, %f794;
|
|
add.f32 %f796, %f792, %f795;
|
|
add.f32 %f797, %f793, %f796;
|
|
sub.f32 %f798, %f793, %f797;
|
|
add.f32 %f799, %f796, %f798;
|
|
mov.f32 %f800, 0f3F317200;
|
|
mul.rn.f32 %f801, %f776, %f800;
|
|
mov.f32 %f802, 0f35BFBE8E;
|
|
mul.rn.f32 %f803, %f776, %f802;
|
|
add.f32 %f804, %f801, %f797;
|
|
sub.f32 %f805, %f801, %f804;
|
|
add.f32 %f806, %f797, %f805;
|
|
add.f32 %f807, %f799, %f806;
|
|
add.f32 %f808, %f803, %f807;
|
|
add.f32 %f809, %f804, %f808;
|
|
sub.f32 %f810, %f804, %f809;
|
|
add.f32 %f811, %f808, %f810;
|
|
mov.f32 %f812, 0f3EE66666;
|
|
mul.rn.f32 %f813, %f812, %f809;
|
|
neg.f32 %f814, %f813;
|
|
fma.rn.f32 %f815, %f812, %f809, %f814;
|
|
fma.rn.f32 %f816, %f812, %f811, %f815;
|
|
mov.f32 %f817, 0f00000000;
|
|
fma.rn.f32 %f818, %f817, %f809, %f816;
|
|
add.rn.f32 %f819, %f813, %f818;
|
|
neg.f32 %f820, %f819;
|
|
add.rn.f32 %f821, %f813, %f820;
|
|
add.rn.f32 %f822, %f821, %f818;
|
|
mov.b32 %r273, %f819;
|
|
setp.eq.s32 %p89, %r273, 1118925336;
|
|
add.s32 %r274, %r273, -1;
|
|
mov.b32 %f823, %r274;
|
|
add.f32 %f824, %f822, 0f37000000;
|
|
selp.f32 %f825, %f823, %f819, %p89;
|
|
selp.f32 %f286, %f824, %f822, %p89;
|
|
mul.f32 %f826, %f825, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f827, %f826;
|
|
mov.f32 %f828, 0fBF317200;
|
|
fma.rn.f32 %f829, %f827, %f828, %f825;
|
|
mov.f32 %f830, 0fB5BFBE8E;
|
|
fma.rn.f32 %f831, %f827, %f830, %f829;
|
|
mul.f32 %f832, %f831, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f833, %f832;
|
|
add.f32 %f834, %f827, 0f00000000;
|
|
ex2.approx.f32 %f835, %f834;
|
|
mul.f32 %f836, %f833, %f835;
|
|
setp.lt.f32 %p90, %f825, 0fC2D20000;
|
|
selp.f32 %f837, 0f00000000, %f836, %p90;
|
|
setp.gt.f32 %p91, %f825, 0f42D20000;
|
|
selp.f32 %f1254, 0f7F800000, %f837, %p91;
|
|
setp.eq.f32 %p92, %f1254, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
fma.rn.f32 %f1254, %f1254, %f286, %f1254;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p93, %f281, 0f00000000;
|
|
setp.eq.f32 %p94, %f284, 0f3F800000;
|
|
and.pred %p4, %p93, %p94;
|
|
mov.b32 %r275, %f1254;
|
|
xor.b32 %r276, %r275, -2147483648;
|
|
mov.b32 %f838, %r276;
|
|
selp.f32 %f1256, %f838, %f1254, %p4;
|
|
setp.eq.f32 %p95, %f281, 0f00000000;
|
|
@%p95 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f841, %f281, %f281;
|
|
selp.f32 %f1256, %f841, 0f00000000, %p94;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p96, %f281, 0f00000000;
|
|
@%p96 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f840, %f812;
|
|
setp.neu.f32 %p97, %f840, 0f3EE66666;
|
|
selp.f32 %f1256, 0f7FFFFFFF, %f1256, %p97;
|
|
|
|
BB0_98:
|
|
add.f32 %f842, %f285, 0f3EE66666;
|
|
mov.b32 %r277, %f842;
|
|
setp.lt.s32 %p99, %r277, 2139095040;
|
|
@%p99 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p100, %f285, 0f7F800000;
|
|
@%p100 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f1256, %f281, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p101, %f285, 0f7F800000;
|
|
@%p101 bra BB0_103;
|
|
|
|
selp.f32 %f1256, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p102, %f281, 0f3F800000;
|
|
selp.f32 %f297, 0f3F800000, %f1256, %p102;
|
|
abs.f32 %f298, %f282;
|
|
setp.lt.f32 %p103, %f298, 0f00800000;
|
|
mul.f32 %f845, %f298, 0f4B800000;
|
|
selp.f32 %f846, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f847, %f845, %f298, %p103;
|
|
mov.b32 %r278, %f847;
|
|
and.b32 %r279, %r278, 8388607;
|
|
or.b32 %r280, %r279, 1065353216;
|
|
mov.b32 %f848, %r280;
|
|
shr.u32 %r281, %r278, 23;
|
|
cvt.rn.f32.u32 %f849, %r281;
|
|
add.f32 %f850, %f846, %f849;
|
|
setp.gt.f32 %p104, %f848, 0f3FB504F3;
|
|
mul.f32 %f851, %f848, 0f3F000000;
|
|
add.f32 %f852, %f850, 0f3F800000;
|
|
selp.f32 %f853, %f851, %f848, %p104;
|
|
selp.f32 %f854, %f852, %f850, %p104;
|
|
add.f32 %f855, %f853, 0fBF800000;
|
|
add.f32 %f844, %f853, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f843,%f844;
|
|
// inline asm
|
|
add.f32 %f856, %f855, %f855;
|
|
mul.f32 %f857, %f843, %f856;
|
|
mul.f32 %f858, %f857, %f857;
|
|
fma.rn.f32 %f861, %f782, %f858, %f781;
|
|
fma.rn.f32 %f863, %f861, %f858, %f784;
|
|
mul.rn.f32 %f864, %f863, %f858;
|
|
mul.rn.f32 %f865, %f864, %f857;
|
|
sub.f32 %f866, %f855, %f857;
|
|
neg.f32 %f867, %f857;
|
|
add.f32 %f868, %f866, %f866;
|
|
fma.rn.f32 %f869, %f867, %f855, %f868;
|
|
mul.rn.f32 %f870, %f843, %f869;
|
|
add.f32 %f871, %f865, %f857;
|
|
sub.f32 %f872, %f857, %f871;
|
|
add.f32 %f873, %f865, %f872;
|
|
add.f32 %f874, %f870, %f873;
|
|
add.f32 %f875, %f871, %f874;
|
|
sub.f32 %f876, %f871, %f875;
|
|
add.f32 %f877, %f874, %f876;
|
|
mul.rn.f32 %f879, %f854, %f800;
|
|
mul.rn.f32 %f881, %f854, %f802;
|
|
add.f32 %f882, %f879, %f875;
|
|
sub.f32 %f883, %f879, %f882;
|
|
add.f32 %f884, %f875, %f883;
|
|
add.f32 %f885, %f877, %f884;
|
|
add.f32 %f886, %f881, %f885;
|
|
add.f32 %f887, %f882, %f886;
|
|
sub.f32 %f888, %f882, %f887;
|
|
add.f32 %f889, %f886, %f888;
|
|
mul.rn.f32 %f891, %f812, %f887;
|
|
neg.f32 %f892, %f891;
|
|
fma.rn.f32 %f893, %f812, %f887, %f892;
|
|
fma.rn.f32 %f894, %f812, %f889, %f893;
|
|
fma.rn.f32 %f896, %f817, %f887, %f894;
|
|
add.rn.f32 %f897, %f891, %f896;
|
|
neg.f32 %f898, %f897;
|
|
add.rn.f32 %f899, %f891, %f898;
|
|
add.rn.f32 %f900, %f899, %f896;
|
|
mov.b32 %r282, %f897;
|
|
setp.eq.s32 %p105, %r282, 1118925336;
|
|
add.s32 %r283, %r282, -1;
|
|
mov.b32 %f901, %r283;
|
|
add.f32 %f902, %f900, 0f37000000;
|
|
selp.f32 %f903, %f901, %f897, %p105;
|
|
selp.f32 %f299, %f902, %f900, %p105;
|
|
mul.f32 %f904, %f903, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f905, %f904;
|
|
fma.rn.f32 %f907, %f905, %f828, %f903;
|
|
fma.rn.f32 %f909, %f905, %f830, %f907;
|
|
mul.f32 %f910, %f909, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f911, %f910;
|
|
add.f32 %f912, %f905, 0f00000000;
|
|
ex2.approx.f32 %f913, %f912;
|
|
mul.f32 %f914, %f911, %f913;
|
|
setp.lt.f32 %p106, %f903, 0fC2D20000;
|
|
selp.f32 %f915, 0f00000000, %f914, %p106;
|
|
setp.gt.f32 %p107, %f903, 0f42D20000;
|
|
selp.f32 %f1257, 0f7F800000, %f915, %p107;
|
|
setp.eq.f32 %p108, %f1257, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
fma.rn.f32 %f1257, %f1257, %f299, %f1257;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p109, %f282, 0f00000000;
|
|
and.pred %p5, %p109, %p94;
|
|
mov.b32 %r284, %f1257;
|
|
xor.b32 %r285, %r284, -2147483648;
|
|
mov.b32 %f916, %r285;
|
|
selp.f32 %f1259, %f916, %f1257, %p5;
|
|
setp.eq.f32 %p111, %f282, 0f00000000;
|
|
@%p111 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f919, %f282, %f282;
|
|
selp.f32 %f1259, %f919, 0f00000000, %p94;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p112, %f282, 0f00000000;
|
|
@%p112 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f918, %f812;
|
|
setp.neu.f32 %p113, %f918, 0f3EE66666;
|
|
selp.f32 %f1259, 0f7FFFFFFF, %f1259, %p113;
|
|
|
|
BB0_109:
|
|
add.f32 %f920, %f298, 0f3EE66666;
|
|
mov.b32 %r286, %f920;
|
|
setp.lt.s32 %p115, %r286, 2139095040;
|
|
@%p115 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p116, %f298, 0f7F800000;
|
|
@%p116 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f1259, %f282, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p117, %f298, 0f7F800000;
|
|
@%p117 bra BB0_114;
|
|
|
|
selp.f32 %f1259, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p118, %f282, 0f3F800000;
|
|
selp.f32 %f310, 0f3F800000, %f1259, %p118;
|
|
abs.f32 %f311, %f283;
|
|
setp.lt.f32 %p119, %f311, 0f00800000;
|
|
mul.f32 %f923, %f311, 0f4B800000;
|
|
selp.f32 %f924, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f925, %f923, %f311, %p119;
|
|
mov.b32 %r287, %f925;
|
|
and.b32 %r288, %r287, 8388607;
|
|
or.b32 %r289, %r288, 1065353216;
|
|
mov.b32 %f926, %r289;
|
|
shr.u32 %r290, %r287, 23;
|
|
cvt.rn.f32.u32 %f927, %r290;
|
|
add.f32 %f928, %f924, %f927;
|
|
setp.gt.f32 %p120, %f926, 0f3FB504F3;
|
|
mul.f32 %f929, %f926, 0f3F000000;
|
|
add.f32 %f930, %f928, 0f3F800000;
|
|
selp.f32 %f931, %f929, %f926, %p120;
|
|
selp.f32 %f932, %f930, %f928, %p120;
|
|
add.f32 %f933, %f931, 0fBF800000;
|
|
add.f32 %f922, %f931, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f921,%f922;
|
|
// inline asm
|
|
add.f32 %f934, %f933, %f933;
|
|
mul.f32 %f935, %f921, %f934;
|
|
mul.f32 %f936, %f935, %f935;
|
|
fma.rn.f32 %f939, %f782, %f936, %f781;
|
|
fma.rn.f32 %f941, %f939, %f936, %f784;
|
|
mul.rn.f32 %f942, %f941, %f936;
|
|
mul.rn.f32 %f943, %f942, %f935;
|
|
sub.f32 %f944, %f933, %f935;
|
|
neg.f32 %f945, %f935;
|
|
add.f32 %f946, %f944, %f944;
|
|
fma.rn.f32 %f947, %f945, %f933, %f946;
|
|
mul.rn.f32 %f948, %f921, %f947;
|
|
add.f32 %f949, %f943, %f935;
|
|
sub.f32 %f950, %f935, %f949;
|
|
add.f32 %f951, %f943, %f950;
|
|
add.f32 %f952, %f948, %f951;
|
|
add.f32 %f953, %f949, %f952;
|
|
sub.f32 %f954, %f949, %f953;
|
|
add.f32 %f955, %f952, %f954;
|
|
mul.rn.f32 %f957, %f932, %f800;
|
|
mul.rn.f32 %f959, %f932, %f802;
|
|
add.f32 %f960, %f957, %f953;
|
|
sub.f32 %f961, %f957, %f960;
|
|
add.f32 %f962, %f953, %f961;
|
|
add.f32 %f963, %f955, %f962;
|
|
add.f32 %f964, %f959, %f963;
|
|
add.f32 %f965, %f960, %f964;
|
|
sub.f32 %f966, %f960, %f965;
|
|
add.f32 %f967, %f964, %f966;
|
|
mul.rn.f32 %f969, %f812, %f965;
|
|
neg.f32 %f970, %f969;
|
|
fma.rn.f32 %f971, %f812, %f965, %f970;
|
|
fma.rn.f32 %f972, %f812, %f967, %f971;
|
|
fma.rn.f32 %f974, %f817, %f965, %f972;
|
|
add.rn.f32 %f975, %f969, %f974;
|
|
neg.f32 %f976, %f975;
|
|
add.rn.f32 %f977, %f969, %f976;
|
|
add.rn.f32 %f978, %f977, %f974;
|
|
mov.b32 %r291, %f975;
|
|
setp.eq.s32 %p121, %r291, 1118925336;
|
|
add.s32 %r292, %r291, -1;
|
|
mov.b32 %f979, %r292;
|
|
add.f32 %f980, %f978, 0f37000000;
|
|
selp.f32 %f981, %f979, %f975, %p121;
|
|
selp.f32 %f312, %f980, %f978, %p121;
|
|
mul.f32 %f982, %f981, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f983, %f982;
|
|
fma.rn.f32 %f985, %f983, %f828, %f981;
|
|
fma.rn.f32 %f987, %f983, %f830, %f985;
|
|
mul.f32 %f988, %f987, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f989, %f988;
|
|
add.f32 %f990, %f983, 0f00000000;
|
|
ex2.approx.f32 %f991, %f990;
|
|
mul.f32 %f992, %f989, %f991;
|
|
setp.lt.f32 %p122, %f981, 0fC2D20000;
|
|
selp.f32 %f993, 0f00000000, %f992, %p122;
|
|
setp.gt.f32 %p123, %f981, 0f42D20000;
|
|
selp.f32 %f1260, 0f7F800000, %f993, %p123;
|
|
setp.eq.f32 %p124, %f1260, 0f7F800000;
|
|
@%p124 bra BB0_116;
|
|
|
|
fma.rn.f32 %f1260, %f1260, %f312, %f1260;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p125, %f283, 0f00000000;
|
|
and.pred %p6, %p125, %p94;
|
|
mov.b32 %r293, %f1260;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f994, %r294;
|
|
selp.f32 %f1262, %f994, %f1260, %p6;
|
|
setp.eq.f32 %p127, %f283, 0f00000000;
|
|
@%p127 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f997, %f283, %f283;
|
|
selp.f32 %f1262, %f997, 0f00000000, %p94;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p128, %f283, 0f00000000;
|
|
@%p128 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f996, %f812;
|
|
setp.neu.f32 %p129, %f996, 0f3EE66666;
|
|
selp.f32 %f1262, 0f7FFFFFFF, %f1262, %p129;
|
|
|
|
BB0_120:
|
|
add.f32 %f998, %f311, 0f3EE66666;
|
|
mov.b32 %r295, %f998;
|
|
setp.lt.s32 %p131, %r295, 2139095040;
|
|
@%p131 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p132, %f311, 0f7F800000;
|
|
@%p132 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f1262, %f283, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p133, %f311, 0f7F800000;
|
|
@%p133 bra BB0_125;
|
|
|
|
selp.f32 %f1262, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
mov.u32 %r382, 4;
|
|
setp.eq.f32 %p134, %f283, 0f3F800000;
|
|
selp.f32 %f999, 0f3F800000, %f1262, %p134;
|
|
cvt.u64.u32 %rd94, %r7;
|
|
cvt.u64.u32 %rd93, %r6;
|
|
mov.u64 %rd97, image;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r101, %r382, %rd93, %rd94, %rd28, %rd28);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1000, %f999;
|
|
mul.f32 %f1001, %f1000, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r298, %f1001;
|
|
cvt.sat.f32.f32 %f1002, %f310;
|
|
mul.f32 %f1003, %f1002, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r299, %f1003;
|
|
cvt.sat.f32.f32 %f1004, %f297;
|
|
mul.f32 %f1005, %f1004, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r300, %f1005;
|
|
cvt.u16.u32 %rs17, %r298;
|
|
cvt.u16.u32 %rs18, %r300;
|
|
cvt.u16.u32 %rs19, %r299;
|
|
mov.u16 %rs20, 255;
|
|
st.v4.u8 [%rd91], {%rs17, %rs19, %rs18, %rs20};
|
|
ld.global.u32 %r411, [imageEnabled];
|
|
|
|
BB0_126:
|
|
cvt.u64.u32 %rd19, %r6;
|
|
cvt.u64.u32 %rd20, %r7;
|
|
and.b32 %r301, %r411, 4;
|
|
setp.eq.s32 %p135, %r301, 0;
|
|
@%p135 bra BB0_130;
|
|
|
|
ld.global.u32 %r302, [additive];
|
|
setp.eq.s32 %p136, %r302, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f549;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r306, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r101, %r306, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs28, %rs29, %rs30, %rs31}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1007, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1008, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1009, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r101, %r306, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1010, %f281, %f1007;
|
|
add.f32 %f1011, %f282, %f1008;
|
|
add.f32 %f1012, %f283, %f1009;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1012;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f1011;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f1010;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs25, %rs26, %rs27, %rs21};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r308, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r101, %r308, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f283;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f282;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f281;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs32, %rs33, %rs34, %rs21};
|
|
|
|
BB0_130:
|
|
selp.f32 %f1017, 0f3F800000, 0f3E800000, %p149;
|
|
mul.f32 %f323, %f1017, %f1244;
|
|
mul.f32 %f324, %f1017, %f1245;
|
|
mul.f32 %f325, %f1017, %f1246;
|
|
mul.f32 %f326, %f1017, %f1241;
|
|
mul.f32 %f327, %f1017, %f1242;
|
|
mul.f32 %f328, %f1017, %f1243;
|
|
mul.f32 %f329, %f1017, %f1238;
|
|
mul.f32 %f330, %f1017, %f1239;
|
|
mul.f32 %f331, %f1017, %f1240;
|
|
mul.f32 %f332, %f1017, %f1235;
|
|
mul.f32 %f333, %f1017, %f1236;
|
|
mul.f32 %f334, %f1017, %f1237;
|
|
ld.global.f32 %f1018, [lightColor];
|
|
mul.f32 %f335, %f323, %f1018;
|
|
ld.global.f32 %f1019, [lightColor+4];
|
|
mul.f32 %f336, %f324, %f1019;
|
|
ld.global.f32 %f1020, [lightColor+8];
|
|
mul.f32 %f337, %f325, %f1020;
|
|
ld.global.u32 %r309, [additive];
|
|
setp.eq.s32 %p137, %r309, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f549;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_132;
|
|
|
|
mov.u64 %rd130, image_RNM0;
|
|
cvta.global.u64 %rd119, %rd130;
|
|
mov.u32 %r313, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r101, %r313, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd118];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1021, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1022, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1023, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd119, %r101, %r313, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1024, %f335, %f1021;
|
|
add.f32 %f1025, %f336, %f1022;
|
|
add.f32 %f1026, %f337, %f1023;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1026;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1025;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f1024;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs39, %rs40, %rs41, %rs35};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd137, image_RNM0;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u32 %r315, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r101, %r315, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f337;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f336;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f335;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs46, %rs47, %rs48, %rs35};
|
|
|
|
BB0_133:
|
|
mov.f32 %f1031, 0f34000000;
|
|
max.f32 %f1032, %f323, %f1031;
|
|
div.rn.f32 %f1033, %f326, %f1032;
|
|
max.f32 %f1034, %f324, %f1031;
|
|
div.rn.f32 %f1035, %f327, %f1034;
|
|
max.f32 %f1036, %f325, %f1031;
|
|
div.rn.f32 %f1037, %f328, %f1036;
|
|
fma.rn.f32 %f338, %f1033, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f339, %f1035, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f340, %f1037, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1038, %f329, %f1032;
|
|
div.rn.f32 %f1039, %f330, %f1034;
|
|
div.rn.f32 %f1040, %f331, %f1036;
|
|
fma.rn.f32 %f341, %f1038, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f342, %f1039, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f343, %f1040, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1041, %f332, %f1032;
|
|
div.rn.f32 %f1042, %f333, %f1034;
|
|
div.rn.f32 %f1043, %f334, %f1036;
|
|
fma.rn.f32 %f344, %f1041, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f345, %f1042, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f346, %f1043, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r316, [additive];
|
|
setp.eq.s32 %p138, %r316, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f549;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_135;
|
|
|
|
mov.u64 %rd150, image_RNM1;
|
|
cvta.global.u64 %rd139, %rd150;
|
|
mov.u32 %r320, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r101, %r320, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd138];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1044, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1045, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1046, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd139, %r101, %r320, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1047, %f338, %f1044;
|
|
add.f32 %f1048, %f339, %f1045;
|
|
add.f32 %f1049, %f340, %f1046;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1049;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1048;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1047;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs53, %rs54, %rs55, %rs49};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd157, image_RNM1;
|
|
cvta.global.u64 %rd152, %rd157;
|
|
mov.u32 %r322, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r101, %r322, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f340;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f339;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f338;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd151], {%rs60, %rs61, %rs62, %rs49};
|
|
|
|
BB0_136:
|
|
ld.global.u32 %r323, [additive];
|
|
setp.eq.s32 %p139, %r323, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f549;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_138;
|
|
|
|
mov.u64 %rd170, image_RNM2;
|
|
cvta.global.u64 %rd159, %rd170;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r101, %r327, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd158];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1054, %rs70;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1055, %rs71;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1056, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd159, %r101, %r327, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1057, %f341, %f1054;
|
|
add.f32 %f1058, %f342, %f1055;
|
|
add.f32 %f1059, %f343, %f1056;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1059;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1058;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f1057;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs67, %rs68, %rs69, %rs63};
|
|
bra.uni BB0_139;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd177, image_RNM2;
|
|
cvta.global.u64 %rd172, %rd177;
|
|
mov.u32 %r329, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r101, %r329, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f343;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f342;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f341;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd171], {%rs74, %rs75, %rs76, %rs63};
|
|
|
|
BB0_139:
|
|
ld.global.u32 %r330, [additive];
|
|
setp.eq.s32 %p140, %r330, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f549;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_141;
|
|
|
|
mov.u64 %rd190, image_RNM3;
|
|
cvta.global.u64 %rd179, %rd190;
|
|
mov.u32 %r334, 8;
|
|
// inline asm
|
|
call (%rd178), _rt_buffer_get_64, (%rd179, %r101, %r334, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd178];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1064, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1065, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1066, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd179, %r101, %r334, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1067, %f344, %f1064;
|
|
add.f32 %f1068, %f345, %f1065;
|
|
add.f32 %f1069, %f346, %f1066;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1069;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f1068;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1067;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_141:
|
|
mov.u64 %rd197, image_RNM3;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u32 %r336, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r101, %r336, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f346;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f345;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f344;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd191], {%rs88, %rs89, %rs90, %rs77};
|
|
|
|
BB0_162:
|
|
ret;
|
|
}
|
|
|
|
|