2315 lines
64 KiB
Plaintext
2315 lines
64 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 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<149>;
|
|
.reg .b16 %rs<152>;
|
|
.reg .f32 %f<988>;
|
|
.reg .b32 %r<411>;
|
|
.reg .b64 %rd<298>;
|
|
|
|
|
|
mov.u64 %rd297, __local_depot0;
|
|
cvta.local.u64 %SP, %rd297;
|
|
ld.global.u32 %r1, [samples];
|
|
shl.b32 %r2, %r1, 1;
|
|
ld.global.v2.u32 {%r100, %r101}, [pixelID];
|
|
cvt.u64.u32 %rd21, %r100;
|
|
cvt.u64.u32 %rd22, %r101;
|
|
mov.u64 %rd25, uvnormal;
|
|
cvta.global.u64 %rd20, %rd25;
|
|
mov.u32 %r98, 2;
|
|
mov.u32 %r99, 4;
|
|
mov.u64 %rd24, 0;
|
|
// inline asm
|
|
call (%rd19), _rt_buffer_get_64, (%rd20, %r98, %r99, %rd21, %rd22, %rd24, %rd24);
|
|
// inline asm
|
|
ld.u32 %r3, [%rd19];
|
|
shr.u32 %r104, %r3, 16;
|
|
cvt.u16.u32 %rs1, %r104;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r3;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p7, %rs9, 0;
|
|
mov.f32 %f916, 0f00000000;
|
|
mov.f32 %f917, %f916;
|
|
mov.f32 %f918, %f916;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd19+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f217, %rs12;
|
|
div.rn.f32 %f218, %f217, 0f437F0000;
|
|
fma.rn.f32 %f219, %f218, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f220, %rs10;
|
|
div.rn.f32 %f221, %f220, 0f437F0000;
|
|
fma.rn.f32 %f222, %f221, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f223, %rs7;
|
|
div.rn.f32 %f224, %f223, 0f437F0000;
|
|
fma.rn.f32 %f225, %f224, 0f40000000, 0fBF800000;
|
|
mul.f32 %f226, %f222, %f222;
|
|
fma.rn.f32 %f227, %f219, %f219, %f226;
|
|
fma.rn.f32 %f228, %f225, %f225, %f227;
|
|
sqrt.rn.f32 %f229, %f228;
|
|
rcp.rn.f32 %f230, %f229;
|
|
mul.f32 %f916, %f219, %f230;
|
|
mul.f32 %f917, %f222, %f230;
|
|
mul.f32 %f918, %f225, %f230;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
ld.global.v2.u32 {%r108, %r109}, [tileInfo];
|
|
add.s32 %r4, %r105, %r108;
|
|
add.s32 %r5, %r106, %r109;
|
|
setp.eq.f32 %p8, %f917, 0f00000000;
|
|
setp.eq.f32 %p9, %f916, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f918, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_139;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_139:
|
|
ld.global.u32 %r410, [imageEnabled];
|
|
and.b32 %r334, %r410, 1;
|
|
setp.eq.b32 %p140, %r334, 1;
|
|
@!%p140 bra BB0_141;
|
|
bra.uni BB0_140;
|
|
|
|
BB0_140:
|
|
cvt.u64.u32 %rd179, %r4;
|
|
cvt.u64.u32 %rd180, %r5;
|
|
mov.u64 %rd183, image;
|
|
cvta.global.u64 %rd178, %rd183;
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r98, %r99, %rd179, %rd180, %rd24, %rd24);
|
|
// inline asm
|
|
mov.u16 %rs84, 0;
|
|
st.v4.u8 [%rd177], {%rs84, %rs84, %rs84, %rs84};
|
|
ld.global.u32 %r410, [imageEnabled];
|
|
|
|
BB0_141:
|
|
and.b32 %r337, %r410, 8;
|
|
setp.eq.s32 %p141, %r337, 0;
|
|
@%p141 bra BB0_143;
|
|
|
|
cvt.u64.u32 %rd187, %r5;
|
|
cvt.u64.u32 %rd186, %r4;
|
|
mov.u64 %rd190, image_Mask;
|
|
cvta.global.u64 %rd185, %rd190;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r98, %r98, %rd186, %rd187, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f873, 0f00000000;
|
|
cvt.rzi.u32.f32 %r340, %f873;
|
|
cvt.u16.u32 %rs85, %r340;
|
|
mov.u16 %rs86, 0;
|
|
st.v2.u8 [%rd184], {%rs85, %rs86};
|
|
ld.global.u32 %r410, [imageEnabled];
|
|
|
|
BB0_143:
|
|
cvt.u64.u32 %rd17, %r4;
|
|
cvt.u64.u32 %rd18, %r5;
|
|
and.b32 %r341, %r410, 4;
|
|
setp.eq.s32 %p142, %r341, 0;
|
|
@%p142 bra BB0_147;
|
|
|
|
ld.global.u32 %r342, [additive];
|
|
setp.eq.s32 %p143, %r342, 0;
|
|
@%p143 bra BB0_146;
|
|
|
|
mov.u64 %rd203, image_HDR;
|
|
cvta.global.u64 %rd192, %rd203;
|
|
mov.u32 %r346, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r98, %r346, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs93, %rs94, %rs95, %rs96}, [%rd191];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f874, %rs93;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f875, %rs94;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f876, %rs95;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd192, %r98, %r346, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f877, %f874, 0f00000000;
|
|
add.f32 %f878, %f875, 0f00000000;
|
|
add.f32 %f879, %f876, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f879;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f878;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f877;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs97, 0;
|
|
st.v4.u16 [%rd197], {%rs90, %rs91, %rs92, %rs97};
|
|
bra.uni BB0_147;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r118, %r119}, [pixelID];
|
|
cvt.u64.u32 %rd28, %r118;
|
|
cvt.u64.u32 %rd29, %r119;
|
|
mov.u64 %rd38, uvpos;
|
|
cvta.global.u64 %rd27, %rd38;
|
|
mov.u32 %r114, 12;
|
|
// inline asm
|
|
call (%rd26), _rt_buffer_get_64, (%rd27, %r98, %r114, %rd28, %rd29, %rd24, %rd24);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd26+8];
|
|
ld.f32 %f8, [%rd26+4];
|
|
ld.f32 %f7, [%rd26];
|
|
mul.f32 %f237, %f7, 0f3456BF95;
|
|
mul.f32 %f238, %f8, 0f3456BF95;
|
|
mul.f32 %f239, %f9, 0f3456BF95;
|
|
abs.f32 %f240, %f916;
|
|
div.rn.f32 %f241, %f237, %f240;
|
|
abs.f32 %f242, %f917;
|
|
div.rn.f32 %f243, %f238, %f242;
|
|
abs.f32 %f244, %f918;
|
|
div.rn.f32 %f245, %f239, %f244;
|
|
abs.f32 %f246, %f241;
|
|
abs.f32 %f247, %f243;
|
|
abs.f32 %f248, %f245;
|
|
mov.f32 %f249, 0f38D1B717;
|
|
max.f32 %f250, %f246, %f249;
|
|
max.f32 %f251, %f247, %f249;
|
|
max.f32 %f252, %f248, %f249;
|
|
fma.rn.f32 %f10, %f916, %f250, %f7;
|
|
fma.rn.f32 %f11, %f917, %f251, %f8;
|
|
fma.rn.f32 %f12, %f918, %f252, %f9;
|
|
ld.global.v2.u32 {%r122, %r123}, [pixelID];
|
|
cvt.u64.u32 %rd34, %r122;
|
|
cvt.u64.u32 %rd35, %r123;
|
|
mov.u64 %rd39, rnd_seeds;
|
|
cvta.global.u64 %rd33, %rd39;
|
|
// inline asm
|
|
call (%rd32), _rt_buffer_get_64, (%rd33, %r98, %r99, %rd34, %rd35, %rd24, %rd24);
|
|
// inline asm
|
|
ld.u32 %r126, [%rd32];
|
|
mad.lo.s32 %r6, %r126, 1664525, 1013904223;
|
|
ld.global.u32 %r127, [lightMeshBufferSize];
|
|
setp.eq.s32 %p14, %r127, 0;
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f18, 0f00000000;
|
|
mov.u32 %r382, 0;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f13, [lightPointSize];
|
|
mul.f32 %f14, %f10, 0f3456BF95;
|
|
mul.f32 %f15, %f11, 0f3456BF95;
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
and.b32 %r129, %r6, 16777215;
|
|
cvt.rn.f32.u32 %f259, %r129;
|
|
mul.f32 %f260, %f259, 0fB3800000;
|
|
fma.rn.f32 %f17, %f260, 0f3F333333, 0f3F800000;
|
|
mov.f32 %f18, 0f00000000;
|
|
mov.u32 %r382, 0;
|
|
abs.f32 %f382, %f15;
|
|
abs.f32 %f383, %f14;
|
|
max.f32 %f384, %f383, %f382;
|
|
abs.f32 %f385, %f16;
|
|
max.f32 %f386, %f384, %f385;
|
|
mov.f32 %f19, %f18;
|
|
mov.f32 %f20, %f18;
|
|
mov.f32 %f21, %f18;
|
|
mov.f32 %f22, %f18;
|
|
mov.f32 %f23, %f18;
|
|
|
|
BB0_6:
|
|
shl.b32 %r132, %r382, 1;
|
|
cvt.s64.s32 %rd42, %r132;
|
|
mov.u64 %rd46, lightMeshBuffer;
|
|
cvta.global.u64 %rd41, %rd46;
|
|
mov.u32 %r130, 1;
|
|
// inline asm
|
|
call (%rd40), _rt_buffer_get_64, (%rd41, %r130, %r114, %rd42, %rd24, %rd24, %rd24);
|
|
// inline asm
|
|
ld.f32 %f261, [%rd40];
|
|
sub.f32 %f262, %f261, %f7;
|
|
ld.f32 %f263, [%rd40+4];
|
|
sub.f32 %f264, %f263, %f8;
|
|
ld.f32 %f265, [%rd40+8];
|
|
sub.f32 %f266, %f265, %f9;
|
|
mul.f32 %f267, %f264, %f264;
|
|
fma.rn.f32 %f268, %f262, %f262, %f267;
|
|
fma.rn.f32 %f269, %f266, %f266, %f268;
|
|
sqrt.rn.f32 %f24, %f269;
|
|
rcp.rn.f32 %f270, %f24;
|
|
mul.f32 %f25, %f262, %f270;
|
|
mul.f32 %f26, %f264, %f270;
|
|
mul.f32 %f27, %f266, %f270;
|
|
mul.f32 %f271, %f24, %f24;
|
|
mul.f32 %f272, %f271, 0f40C90FDB;
|
|
div.rn.f32 %f273, %f13, %f272;
|
|
add.f32 %f28, %f273, %f273;
|
|
setp.gt.f32 %p16, %f28, %f17;
|
|
setp.ne.s32 %p17, %r1, 0;
|
|
and.pred %p18, %p17, %p16;
|
|
mov.pred %p148, -1;
|
|
@%p18 bra BB0_22;
|
|
|
|
ld.global.f32 %f276, [lightInvCutoff];
|
|
mul.f32 %f29, %f24, %f276;
|
|
mov.f32 %f280, 0f40800000;
|
|
abs.f32 %f31, %f29;
|
|
setp.lt.f32 %p19, %f31, 0f00800000;
|
|
mul.f32 %f282, %f31, 0f4B800000;
|
|
selp.f32 %f283, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f284, %f282, %f31, %p19;
|
|
mov.b32 %r133, %f284;
|
|
and.b32 %r134, %r133, 8388607;
|
|
or.b32 %r135, %r134, 1065353216;
|
|
mov.b32 %f285, %r135;
|
|
shr.u32 %r136, %r133, 23;
|
|
cvt.rn.f32.u32 %f286, %r136;
|
|
add.f32 %f287, %f283, %f286;
|
|
setp.gt.f32 %p20, %f285, 0f3FB504F3;
|
|
mul.f32 %f288, %f285, 0f3F000000;
|
|
add.f32 %f289, %f287, 0f3F800000;
|
|
selp.f32 %f290, %f288, %f285, %p20;
|
|
selp.f32 %f291, %f289, %f287, %p20;
|
|
add.f32 %f292, %f290, 0fBF800000;
|
|
add.f32 %f275, %f290, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f274,%f275;
|
|
// inline asm
|
|
add.f32 %f293, %f292, %f292;
|
|
mul.f32 %f294, %f274, %f293;
|
|
mul.f32 %f295, %f294, %f294;
|
|
mov.f32 %f296, 0f3C4CAF63;
|
|
mov.f32 %f297, 0f3B18F0FE;
|
|
fma.rn.f32 %f298, %f297, %f295, %f296;
|
|
mov.f32 %f299, 0f3DAAAABD;
|
|
fma.rn.f32 %f300, %f298, %f295, %f299;
|
|
mul.rn.f32 %f301, %f300, %f295;
|
|
mul.rn.f32 %f302, %f301, %f294;
|
|
sub.f32 %f303, %f292, %f294;
|
|
neg.f32 %f304, %f294;
|
|
add.f32 %f305, %f303, %f303;
|
|
fma.rn.f32 %f306, %f304, %f292, %f305;
|
|
mul.rn.f32 %f307, %f274, %f306;
|
|
add.f32 %f308, %f302, %f294;
|
|
sub.f32 %f309, %f294, %f308;
|
|
add.f32 %f310, %f302, %f309;
|
|
add.f32 %f311, %f307, %f310;
|
|
add.f32 %f312, %f308, %f311;
|
|
sub.f32 %f313, %f308, %f312;
|
|
add.f32 %f314, %f311, %f313;
|
|
mov.f32 %f315, 0f3F317200;
|
|
mul.rn.f32 %f316, %f291, %f315;
|
|
mov.f32 %f317, 0f35BFBE8E;
|
|
mul.rn.f32 %f318, %f291, %f317;
|
|
add.f32 %f319, %f316, %f312;
|
|
sub.f32 %f320, %f316, %f319;
|
|
add.f32 %f321, %f312, %f320;
|
|
add.f32 %f322, %f314, %f321;
|
|
add.f32 %f323, %f318, %f322;
|
|
add.f32 %f324, %f319, %f323;
|
|
sub.f32 %f325, %f319, %f324;
|
|
add.f32 %f326, %f323, %f325;
|
|
mul.rn.f32 %f327, %f280, %f324;
|
|
neg.f32 %f328, %f327;
|
|
fma.rn.f32 %f329, %f280, %f324, %f328;
|
|
fma.rn.f32 %f330, %f280, %f326, %f329;
|
|
mov.f32 %f331, 0f00000000;
|
|
fma.rn.f32 %f332, %f331, %f324, %f330;
|
|
add.rn.f32 %f333, %f327, %f332;
|
|
neg.f32 %f334, %f333;
|
|
add.rn.f32 %f335, %f327, %f334;
|
|
add.rn.f32 %f336, %f335, %f332;
|
|
mov.b32 %r137, %f333;
|
|
setp.eq.s32 %p21, %r137, 1118925336;
|
|
add.s32 %r138, %r137, -1;
|
|
mov.b32 %f337, %r138;
|
|
add.f32 %f338, %f336, 0f37000000;
|
|
selp.f32 %f339, %f337, %f333, %p21;
|
|
selp.f32 %f32, %f338, %f336, %p21;
|
|
mul.f32 %f340, %f339, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f341, %f340;
|
|
mov.f32 %f342, 0fBF317200;
|
|
fma.rn.f32 %f343, %f341, %f342, %f339;
|
|
mov.f32 %f344, 0fB5BFBE8E;
|
|
fma.rn.f32 %f345, %f341, %f344, %f343;
|
|
mul.f32 %f346, %f345, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f347, %f346;
|
|
add.f32 %f348, %f341, 0f00000000;
|
|
ex2.approx.f32 %f349, %f348;
|
|
mul.f32 %f350, %f347, %f349;
|
|
setp.lt.f32 %p22, %f339, 0fC2D20000;
|
|
selp.f32 %f351, 0f00000000, %f350, %p22;
|
|
setp.gt.f32 %p23, %f339, 0f42D20000;
|
|
selp.f32 %f925, 0f7F800000, %f351, %p23;
|
|
setp.eq.f32 %p24, %f925, 0f7F800000;
|
|
@%p24 bra BB0_9;
|
|
|
|
fma.rn.f32 %f925, %f925, %f32, %f925;
|
|
|
|
BB0_9:
|
|
mov.f32 %f914, 0f40000000;
|
|
cvt.rzi.f32.f32 %f913, %f914;
|
|
add.f32 %f912, %f913, %f913;
|
|
mov.f32 %f911, 0f40800000;
|
|
sub.f32 %f910, %f911, %f912;
|
|
abs.f32 %f909, %f910;
|
|
setp.lt.f32 %p25, %f29, 0f00000000;
|
|
setp.eq.f32 %p26, %f909, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r139, %f925;
|
|
xor.b32 %r140, %r139, -2147483648;
|
|
mov.b32 %f352, %r140;
|
|
selp.f32 %f927, %f352, %f925, %p1;
|
|
setp.eq.f32 %p27, %f29, 0f00000000;
|
|
@%p27 bra BB0_12;
|
|
bra.uni BB0_10;
|
|
|
|
BB0_12:
|
|
add.f32 %f355, %f29, %f29;
|
|
selp.f32 %f927, %f355, 0f00000000, %p26;
|
|
bra.uni BB0_13;
|
|
|
|
BB0_10:
|
|
setp.geu.f32 %p28, %f29, 0f00000000;
|
|
@%p28 bra BB0_13;
|
|
|
|
mov.f32 %f915, 0f40800000;
|
|
cvt.rzi.f32.f32 %f354, %f915;
|
|
setp.neu.f32 %p29, %f354, 0f40800000;
|
|
selp.f32 %f927, 0f7FFFFFFF, %f927, %p29;
|
|
|
|
BB0_13:
|
|
add.f32 %f356, %f31, 0f40800000;
|
|
mov.b32 %r141, %f356;
|
|
setp.lt.s32 %p31, %r141, 2139095040;
|
|
@%p31 bra BB0_18;
|
|
|
|
setp.gtu.f32 %p32, %f31, 0f7F800000;
|
|
@%p32 bra BB0_17;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_17:
|
|
add.f32 %f927, %f29, 0f40800000;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_15:
|
|
setp.neu.f32 %p33, %f31, 0f7F800000;
|
|
@%p33 bra BB0_18;
|
|
|
|
selp.f32 %f927, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_18:
|
|
mov.u32 %r378, 1;
|
|
mov.u64 %rd292, lightMeshBuffer;
|
|
cvta.global.u64 %rd291, %rd292;
|
|
shl.b32 %r377, %r382, 1;
|
|
mov.f32 %f357, 0f3F800000;
|
|
sub.f32 %f358, %f357, %f927;
|
|
setp.eq.f32 %p34, %f29, 0f3F800000;
|
|
selp.f32 %f359, 0f00000000, %f358, %p34;
|
|
cvt.sat.f32.f32 %f360, %f359;
|
|
mul.f32 %f361, %f28, %f360;
|
|
add.s32 %r145, %r377, 1;
|
|
cvt.s64.s32 %rd49, %r145;
|
|
// inline asm
|
|
call (%rd47), _rt_buffer_get_64, (%rd291, %r378, %r114, %rd49, %rd24, %rd24, %rd24);
|
|
// inline asm
|
|
ld.f32 %f362, [%rd47];
|
|
mul.f32 %f363, %f25, %f362;
|
|
ld.f32 %f364, [%rd47+4];
|
|
mul.f32 %f365, %f26, %f364;
|
|
neg.f32 %f366, %f365;
|
|
sub.f32 %f367, %f366, %f363;
|
|
ld.f32 %f368, [%rd47+8];
|
|
mul.f32 %f369, %f27, %f368;
|
|
sub.f32 %f370, %f367, %f369;
|
|
cvt.sat.f32.f32 %f371, %f370;
|
|
mul.f32 %f43, %f361, %f371;
|
|
mul.f32 %f372, %f917, %f26;
|
|
fma.rn.f32 %f373, %f916, %f25, %f372;
|
|
fma.rn.f32 %f44, %f918, %f27, %f373;
|
|
setp.leu.f32 %p35, %f43, 0f3727C5AC;
|
|
@%p35 bra BB0_20;
|
|
|
|
mov.u32 %r380, 1;
|
|
add.u64 %rd54, %SP, 4;
|
|
cvta.to.local.u64 %rd55, %rd54;
|
|
max.f32 %f380, %f386, %f249;
|
|
sub.f32 %f381, %f24, %f380;
|
|
mov.u32 %r149, 1065353216;
|
|
st.local.u32 [%rd55], %r149;
|
|
ld.global.u32 %r146, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r146, %f10, %f11, %f12, %f25, %f26, %f27, %r380, %f380, %f381, %rd54, %r99);
|
|
// inline asm
|
|
ld.local.f32 %f388, [%rd55];
|
|
mul.f32 %f389, %f43, %f388;
|
|
cvt.sat.f32.f32 %f390, %f44;
|
|
fma.rn.f32 %f23, %f390, %f389, %f23;
|
|
fma.rn.f32 %f21, %f25, %f389, %f21;
|
|
fma.rn.f32 %f20, %f26, %f389, %f20;
|
|
fma.rn.f32 %f19, %f27, %f389, %f19;
|
|
add.f32 %f22, %f22, %f389;
|
|
add.f32 %f18, %f18, %f388;
|
|
|
|
BB0_20:
|
|
ld.global.u32 %r150, [lightMeshBufferSize];
|
|
add.s32 %r382, %r382, 1;
|
|
setp.lt.u32 %p37, %r382, %r150;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_4:
|
|
mov.f32 %f19, %f18;
|
|
mov.f32 %f20, %f18;
|
|
mov.f32 %f21, %f18;
|
|
mov.f32 %f22, %f18;
|
|
mov.f32 %f23, %f18;
|
|
|
|
BB0_21:
|
|
mov.pred %p148, %p13;
|
|
|
|
BB0_22:
|
|
cvt.rn.f32.s32 %f391, %r382;
|
|
mov.f32 %f392, 0f3F800000;
|
|
max.f32 %f393, %f391, %f392;
|
|
div.rn.f32 %f970, %f23, %f393;
|
|
div.rn.f32 %f975, %f18, %f393;
|
|
div.rn.f32 %f971, %f22, %f393;
|
|
div.rn.f32 %f972, %f21, %f393;
|
|
div.rn.f32 %f973, %f20, %f393;
|
|
div.rn.f32 %f974, %f19, %f393;
|
|
@!%p148 bra BB0_75;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_23:
|
|
mov.f32 %f946, 0f00000000;
|
|
setp.lt.s32 %p38, %r1, 1;
|
|
mov.f32 %f947, %f946;
|
|
mov.f32 %f948, %f946;
|
|
mov.f32 %f949, %f946;
|
|
mov.f32 %f950, %f946;
|
|
mov.f32 %f951, %f946;
|
|
@%p38 bra BB0_74;
|
|
|
|
mad.lo.s32 %r386, %r126, 1664525, 1013904223;
|
|
cvt.rn.f32.s32 %f406, %r2;
|
|
rcp.rn.f32 %f69, %f406;
|
|
add.u64 %rd56, %SP, 8;
|
|
cvta.to.local.u64 %rd2, %rd56;
|
|
mul.f32 %f70, %f10, 0f3456BF95;
|
|
mul.f32 %f71, %f11, 0f3456BF95;
|
|
mul.f32 %f72, %f12, 0f3456BF95;
|
|
add.u64 %rd57, %SP, 0;
|
|
cvta.to.local.u64 %rd3, %rd57;
|
|
mov.f32 %f946, 0f00000000;
|
|
mov.u32 %r151, 0;
|
|
abs.f32 %f407, %f71;
|
|
abs.f32 %f408, %f70;
|
|
max.f32 %f409, %f408, %f407;
|
|
abs.f32 %f410, %f72;
|
|
max.f32 %f411, %f409, %f410;
|
|
mov.u32 %r383, %r151;
|
|
mov.f32 %f947, %f946;
|
|
mov.f32 %f948, %f946;
|
|
mov.f32 %f949, %f946;
|
|
mov.f32 %f950, %f946;
|
|
mov.f32 %f951, %f946;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f79, %r383;
|
|
max.f32 %f80, %f411, %f249;
|
|
mov.u32 %r385, %r151;
|
|
|
|
BB0_26:
|
|
mad.lo.s32 %r153, %r386, 1664525, 1013904223;
|
|
and.b32 %r154, %r153, 16777215;
|
|
cvt.rn.f32.u32 %f413, %r154;
|
|
fma.rn.f32 %f414, %f413, 0f33800000, %f79;
|
|
mul.f32 %f415, %f69, %f414;
|
|
mad.lo.s32 %r386, %r153, 1664525, 1013904223;
|
|
and.b32 %r155, %r386, 16777215;
|
|
cvt.rn.f32.u32 %f416, %r155;
|
|
cvt.rn.f32.s32 %f417, %r385;
|
|
fma.rn.f32 %f418, %f416, 0f33800000, %f417;
|
|
mul.f32 %f419, %f69, %f418;
|
|
fma.rn.f32 %f87, %f415, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f420, %f87, %f87;
|
|
sub.f32 %f422, %f392, %f420;
|
|
mov.f32 %f423, 0f00000000;
|
|
max.f32 %f424, %f423, %f422;
|
|
sqrt.rn.f32 %f88, %f424;
|
|
mul.f32 %f958, %f419, 0f40C90FDB;
|
|
abs.f32 %f90, %f958;
|
|
setp.neu.f32 %p39, %f90, 0f7F800000;
|
|
mov.f32 %f952, %f958;
|
|
@%p39 bra BB0_28;
|
|
|
|
mul.rn.f32 %f952, %f958, %f423;
|
|
|
|
BB0_28:
|
|
mul.f32 %f426, %f952, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r396, %f426;
|
|
cvt.rn.f32.s32 %f427, %r396;
|
|
neg.f32 %f428, %f427;
|
|
mov.f32 %f429, 0f3FC90FDA;
|
|
fma.rn.f32 %f430, %f428, %f429, %f952;
|
|
mov.f32 %f431, 0f33A22168;
|
|
fma.rn.f32 %f432, %f428, %f431, %f430;
|
|
mov.f32 %f433, 0f27C234C5;
|
|
fma.rn.f32 %f953, %f428, %f433, %f432;
|
|
abs.f32 %f434, %f952;
|
|
setp.leu.f32 %p40, %f434, 0f47CE4780;
|
|
@%p40 bra BB0_39;
|
|
|
|
mov.b32 %r16, %f952;
|
|
shr.u32 %r17, %r16, 23;
|
|
shl.b32 %r158, %r16, 8;
|
|
or.b32 %r18, %r158, -2147483648;
|
|
mov.u32 %r388, 0;
|
|
mov.u64 %rd293, __cudart_i2opi_f;
|
|
mov.u32 %r387, -6;
|
|
mov.u64 %rd294, %rd2;
|
|
|
|
BB0_30:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r161, [%rd293];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r159, %r161, %r18, %r388;
|
|
madc.hi.u32 %r388, %r161, %r18, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd294], %r159;
|
|
add.s64 %rd294, %rd294, 4;
|
|
add.s64 %rd293, %rd293, 4;
|
|
add.s32 %r387, %r387, 1;
|
|
setp.ne.s32 %p41, %r387, 0;
|
|
@%p41 bra BB0_30;
|
|
|
|
and.b32 %r164, %r17, 255;
|
|
add.s32 %r165, %r164, -128;
|
|
shr.u32 %r166, %r165, 5;
|
|
and.b32 %r23, %r16, -2147483648;
|
|
st.local.u32 [%rd2+24], %r388;
|
|
mov.u32 %r167, 6;
|
|
sub.s32 %r168, %r167, %r166;
|
|
mul.wide.s32 %rd59, %r168, 4;
|
|
add.s64 %rd8, %rd2, %rd59;
|
|
ld.local.u32 %r389, [%rd8];
|
|
ld.local.u32 %r390, [%rd8+-4];
|
|
and.b32 %r26, %r17, 31;
|
|
setp.eq.s32 %p42, %r26, 0;
|
|
@%p42 bra BB0_33;
|
|
|
|
mov.u32 %r169, 32;
|
|
sub.s32 %r170, %r169, %r26;
|
|
shr.u32 %r171, %r390, %r170;
|
|
shl.b32 %r172, %r389, %r26;
|
|
add.s32 %r389, %r171, %r172;
|
|
ld.local.u32 %r173, [%rd8+-8];
|
|
shr.u32 %r174, %r173, %r170;
|
|
shl.b32 %r175, %r390, %r26;
|
|
add.s32 %r390, %r174, %r175;
|
|
|
|
BB0_33:
|
|
shr.u32 %r176, %r390, 30;
|
|
shl.b32 %r177, %r389, 2;
|
|
add.s32 %r391, %r176, %r177;
|
|
shl.b32 %r32, %r390, 2;
|
|
shr.u32 %r178, %r391, 31;
|
|
shr.u32 %r179, %r389, 30;
|
|
add.s32 %r33, %r178, %r179;
|
|
setp.eq.s32 %p43, %r178, 0;
|
|
@%p43 bra BB0_34;
|
|
bra.uni BB0_35;
|
|
|
|
BB0_34:
|
|
mov.u32 %r392, %r23;
|
|
mov.u32 %r393, %r32;
|
|
bra.uni BB0_36;
|
|
|
|
BB0_35:
|
|
not.b32 %r180, %r391;
|
|
neg.s32 %r393, %r32;
|
|
setp.eq.s32 %p44, %r32, 0;
|
|
selp.u32 %r181, 1, 0, %p44;
|
|
add.s32 %r391, %r181, %r180;
|
|
xor.b32 %r392, %r23, -2147483648;
|
|
|
|
BB0_36:
|
|
clz.b32 %r395, %r391;
|
|
setp.eq.s32 %p45, %r395, 0;
|
|
shl.b32 %r182, %r391, %r395;
|
|
mov.u32 %r183, 32;
|
|
sub.s32 %r184, %r183, %r395;
|
|
shr.u32 %r185, %r393, %r184;
|
|
add.s32 %r186, %r185, %r182;
|
|
selp.b32 %r41, %r391, %r186, %p45;
|
|
mov.u32 %r187, -921707870;
|
|
mul.hi.u32 %r394, %r41, %r187;
|
|
setp.eq.s32 %p46, %r23, 0;
|
|
neg.s32 %r188, %r33;
|
|
selp.b32 %r396, %r33, %r188, %p46;
|
|
setp.lt.s32 %p47, %r394, 1;
|
|
@%p47 bra BB0_38;
|
|
|
|
mul.lo.s32 %r189, %r41, -921707870;
|
|
shr.u32 %r190, %r189, 31;
|
|
shl.b32 %r191, %r394, 1;
|
|
add.s32 %r394, %r190, %r191;
|
|
add.s32 %r395, %r395, 1;
|
|
|
|
BB0_38:
|
|
mov.u32 %r192, 126;
|
|
sub.s32 %r193, %r192, %r395;
|
|
shl.b32 %r194, %r193, 23;
|
|
add.s32 %r195, %r394, 1;
|
|
shr.u32 %r196, %r195, 7;
|
|
add.s32 %r197, %r196, 1;
|
|
shr.u32 %r198, %r197, 1;
|
|
add.s32 %r199, %r198, %r194;
|
|
or.b32 %r200, %r199, %r392;
|
|
mov.b32 %f953, %r200;
|
|
|
|
BB0_39:
|
|
mul.rn.f32 %f96, %f953, %f953;
|
|
add.s32 %r49, %r396, 1;
|
|
and.b32 %r50, %r49, 1;
|
|
setp.eq.s32 %p48, %r50, 0;
|
|
@%p48 bra BB0_41;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_41:
|
|
mov.f32 %f437, 0f3C08839E;
|
|
mov.f32 %f438, 0fB94CA1F9;
|
|
fma.rn.f32 %f954, %f438, %f96, %f437;
|
|
bra.uni BB0_42;
|
|
|
|
BB0_40:
|
|
mov.f32 %f435, 0fBAB6061A;
|
|
mov.f32 %f436, 0f37CCF5CE;
|
|
fma.rn.f32 %f954, %f436, %f96, %f435;
|
|
|
|
BB0_42:
|
|
@%p48 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f442, 0fBE2AAAA3;
|
|
fma.rn.f32 %f443, %f954, %f96, %f442;
|
|
fma.rn.f32 %f955, %f443, %f96, %f423;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f439, 0f3D2AAAA5;
|
|
fma.rn.f32 %f440, %f954, %f96, %f439;
|
|
mov.f32 %f441, 0fBF000000;
|
|
fma.rn.f32 %f955, %f440, %f96, %f441;
|
|
|
|
BB0_45:
|
|
fma.rn.f32 %f956, %f955, %f953, %f953;
|
|
@%p48 bra BB0_47;
|
|
|
|
fma.rn.f32 %f956, %f955, %f96, %f392;
|
|
|
|
BB0_47:
|
|
and.b32 %r201, %r49, 2;
|
|
setp.eq.s32 %p51, %r201, 0;
|
|
@%p51 bra BB0_49;
|
|
|
|
mov.f32 %f447, 0fBF800000;
|
|
fma.rn.f32 %f956, %f956, %f447, %f423;
|
|
|
|
BB0_49:
|
|
@%p39 bra BB0_51;
|
|
|
|
mul.rn.f32 %f958, %f958, %f423;
|
|
|
|
BB0_51:
|
|
mul.f32 %f449, %f958, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r406, %f449;
|
|
cvt.rn.f32.s32 %f450, %r406;
|
|
neg.f32 %f451, %f450;
|
|
fma.rn.f32 %f453, %f451, %f429, %f958;
|
|
fma.rn.f32 %f455, %f451, %f431, %f453;
|
|
fma.rn.f32 %f959, %f451, %f433, %f455;
|
|
abs.f32 %f457, %f958;
|
|
setp.leu.f32 %p53, %f457, 0f47CE4780;
|
|
@%p53 bra BB0_62;
|
|
|
|
mov.b32 %r52, %f958;
|
|
shr.u32 %r53, %r52, 23;
|
|
shl.b32 %r204, %r52, 8;
|
|
or.b32 %r54, %r204, -2147483648;
|
|
mov.u32 %r398, 0;
|
|
mov.u64 %rd295, __cudart_i2opi_f;
|
|
mov.u32 %r397, -6;
|
|
mov.u64 %rd296, %rd2;
|
|
|
|
BB0_53:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r207, [%rd295];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r205, %r207, %r54, %r398;
|
|
madc.hi.u32 %r398, %r207, %r54, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd296], %r205;
|
|
add.s64 %rd296, %rd296, 4;
|
|
add.s64 %rd295, %rd295, 4;
|
|
add.s32 %r397, %r397, 1;
|
|
setp.ne.s32 %p54, %r397, 0;
|
|
@%p54 bra BB0_53;
|
|
|
|
and.b32 %r210, %r53, 255;
|
|
add.s32 %r211, %r210, -128;
|
|
shr.u32 %r212, %r211, 5;
|
|
and.b32 %r59, %r52, -2147483648;
|
|
st.local.u32 [%rd2+24], %r398;
|
|
mov.u32 %r213, 6;
|
|
sub.s32 %r214, %r213, %r212;
|
|
mul.wide.s32 %rd61, %r214, 4;
|
|
add.s64 %rd13, %rd2, %rd61;
|
|
ld.local.u32 %r399, [%rd13];
|
|
ld.local.u32 %r400, [%rd13+-4];
|
|
and.b32 %r62, %r53, 31;
|
|
setp.eq.s32 %p55, %r62, 0;
|
|
@%p55 bra BB0_56;
|
|
|
|
mov.u32 %r215, 32;
|
|
sub.s32 %r216, %r215, %r62;
|
|
shr.u32 %r217, %r400, %r216;
|
|
shl.b32 %r218, %r399, %r62;
|
|
add.s32 %r399, %r217, %r218;
|
|
ld.local.u32 %r219, [%rd13+-8];
|
|
shr.u32 %r220, %r219, %r216;
|
|
shl.b32 %r221, %r400, %r62;
|
|
add.s32 %r400, %r220, %r221;
|
|
|
|
BB0_56:
|
|
shr.u32 %r222, %r400, 30;
|
|
shl.b32 %r223, %r399, 2;
|
|
add.s32 %r401, %r222, %r223;
|
|
shl.b32 %r68, %r400, 2;
|
|
shr.u32 %r224, %r401, 31;
|
|
shr.u32 %r225, %r399, 30;
|
|
add.s32 %r69, %r224, %r225;
|
|
setp.eq.s32 %p56, %r224, 0;
|
|
@%p56 bra BB0_57;
|
|
bra.uni BB0_58;
|
|
|
|
BB0_57:
|
|
mov.u32 %r402, %r59;
|
|
mov.u32 %r403, %r68;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_58:
|
|
not.b32 %r226, %r401;
|
|
neg.s32 %r403, %r68;
|
|
setp.eq.s32 %p57, %r68, 0;
|
|
selp.u32 %r227, 1, 0, %p57;
|
|
add.s32 %r401, %r227, %r226;
|
|
xor.b32 %r402, %r59, -2147483648;
|
|
|
|
BB0_59:
|
|
clz.b32 %r405, %r401;
|
|
setp.eq.s32 %p58, %r405, 0;
|
|
shl.b32 %r228, %r401, %r405;
|
|
mov.u32 %r229, 32;
|
|
sub.s32 %r230, %r229, %r405;
|
|
shr.u32 %r231, %r403, %r230;
|
|
add.s32 %r232, %r231, %r228;
|
|
selp.b32 %r77, %r401, %r232, %p58;
|
|
mov.u32 %r233, -921707870;
|
|
mul.hi.u32 %r404, %r77, %r233;
|
|
setp.eq.s32 %p59, %r59, 0;
|
|
neg.s32 %r234, %r69;
|
|
selp.b32 %r406, %r69, %r234, %p59;
|
|
setp.lt.s32 %p60, %r404, 1;
|
|
@%p60 bra BB0_61;
|
|
|
|
mul.lo.s32 %r235, %r77, -921707870;
|
|
shr.u32 %r236, %r235, 31;
|
|
shl.b32 %r237, %r404, 1;
|
|
add.s32 %r404, %r236, %r237;
|
|
add.s32 %r405, %r405, 1;
|
|
|
|
BB0_61:
|
|
mov.u32 %r238, 126;
|
|
sub.s32 %r239, %r238, %r405;
|
|
shl.b32 %r240, %r239, 23;
|
|
add.s32 %r241, %r404, 1;
|
|
shr.u32 %r242, %r241, 7;
|
|
add.s32 %r243, %r242, 1;
|
|
shr.u32 %r244, %r243, 1;
|
|
add.s32 %r245, %r244, %r240;
|
|
or.b32 %r246, %r245, %r402;
|
|
mov.b32 %f959, %r246;
|
|
|
|
BB0_62:
|
|
mul.rn.f32 %f113, %f959, %f959;
|
|
and.b32 %r85, %r406, 1;
|
|
setp.eq.s32 %p61, %r85, 0;
|
|
@%p61 bra BB0_64;
|
|
bra.uni BB0_63;
|
|
|
|
BB0_64:
|
|
mov.f32 %f460, 0f3C08839E;
|
|
mov.f32 %f461, 0fB94CA1F9;
|
|
fma.rn.f32 %f960, %f461, %f113, %f460;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_63:
|
|
mov.f32 %f458, 0fBAB6061A;
|
|
mov.f32 %f459, 0f37CCF5CE;
|
|
fma.rn.f32 %f960, %f459, %f113, %f458;
|
|
|
|
BB0_65:
|
|
@%p61 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f465, 0fBE2AAAA3;
|
|
fma.rn.f32 %f466, %f960, %f113, %f465;
|
|
fma.rn.f32 %f961, %f466, %f113, %f423;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f462, 0f3D2AAAA5;
|
|
fma.rn.f32 %f463, %f960, %f113, %f462;
|
|
mov.f32 %f464, 0fBF000000;
|
|
fma.rn.f32 %f961, %f463, %f113, %f464;
|
|
|
|
BB0_68:
|
|
fma.rn.f32 %f962, %f961, %f959, %f959;
|
|
@%p61 bra BB0_70;
|
|
|
|
fma.rn.f32 %f962, %f961, %f113, %f392;
|
|
|
|
BB0_70:
|
|
and.b32 %r247, %r406, 2;
|
|
setp.eq.s32 %p64, %r247, 0;
|
|
@%p64 bra BB0_72;
|
|
|
|
mov.f32 %f470, 0fBF800000;
|
|
fma.rn.f32 %f962, %f962, %f470, %f423;
|
|
|
|
BB0_72:
|
|
mul.f32 %f474, %f88, %f956;
|
|
mov.u32 %r249, 0;
|
|
st.local.u32 [%rd3], %r249;
|
|
ld.global.u32 %r248, [root];
|
|
mul.f32 %f475, %f88, %f962;
|
|
mov.f32 %f478, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r248, %f10, %f11, %f12, %f474, %f475, %f87, %r249, %f80, %f478, %rd57, %r99);
|
|
// inline asm
|
|
ld.local.f32 %f479, [%rd3];
|
|
setp.lt.f32 %p65, %f479, 0f00000000;
|
|
selp.f32 %f480, 0f00000000, %f479, %p65;
|
|
selp.f32 %f481, 0f00000000, 0f3F800000, %p65;
|
|
add.f32 %f946, %f946, %f481;
|
|
fma.rn.f32 %f949, %f474, %f480, %f949;
|
|
fma.rn.f32 %f948, %f475, %f480, %f948;
|
|
fma.rn.f32 %f947, %f87, %f480, %f947;
|
|
add.f32 %f950, %f950, %f480;
|
|
mul.f32 %f482, %f917, %f475;
|
|
fma.rn.f32 %f483, %f916, %f474, %f482;
|
|
fma.rn.f32 %f484, %f918, %f87, %f483;
|
|
cvt.sat.f32.f32 %f485, %f484;
|
|
fma.rn.f32 %f951, %f480, %f485, %f951;
|
|
add.s32 %r385, %r385, 1;
|
|
setp.lt.s32 %p66, %r385, %r2;
|
|
@%p66 bra BB0_26;
|
|
|
|
add.s32 %r383, %r383, 1;
|
|
setp.lt.s32 %p67, %r383, %r2;
|
|
@%p67 bra BB0_25;
|
|
|
|
BB0_74:
|
|
mul.lo.s32 %r251, %r2, %r2;
|
|
cvt.rn.f32.s32 %f486, %r251;
|
|
div.rn.f32 %f487, %f951, %f486;
|
|
div.rn.f32 %f975, %f946, %f486;
|
|
div.rn.f32 %f971, %f950, %f486;
|
|
div.rn.f32 %f972, %f949, %f486;
|
|
div.rn.f32 %f973, %f948, %f486;
|
|
div.rn.f32 %f974, %f947, %f486;
|
|
add.f32 %f970, %f487, %f487;
|
|
|
|
BB0_75:
|
|
ld.global.u32 %r408, [imageEnabled];
|
|
and.b32 %r252, %r408, 8;
|
|
setp.eq.s32 %p68, %r252, 0;
|
|
@%p68 bra BB0_88;
|
|
|
|
cvt.u64.u32 %rd65, %r4;
|
|
cvt.u64.u32 %rd66, %r5;
|
|
mov.u64 %rd69, image_Mask;
|
|
cvta.global.u64 %rd64, %rd69;
|
|
// inline asm
|
|
call (%rd63), _rt_buffer_get_64, (%rd64, %r98, %r98, %rd65, %rd66, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f490, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f491, %f490;
|
|
fma.rn.f32 %f492, %f491, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f149, %f492;
|
|
abs.f32 %f150, %f975;
|
|
setp.lt.f32 %p69, %f150, 0f00800000;
|
|
mul.f32 %f493, %f150, 0f4B800000;
|
|
selp.f32 %f494, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f495, %f493, %f150, %p69;
|
|
mov.b32 %r255, %f495;
|
|
and.b32 %r256, %r255, 8388607;
|
|
or.b32 %r257, %r256, 1065353216;
|
|
mov.b32 %f496, %r257;
|
|
shr.u32 %r258, %r255, 23;
|
|
cvt.rn.f32.u32 %f497, %r258;
|
|
add.f32 %f498, %f494, %f497;
|
|
setp.gt.f32 %p70, %f496, 0f3FB504F3;
|
|
mul.f32 %f499, %f496, 0f3F000000;
|
|
add.f32 %f500, %f498, 0f3F800000;
|
|
selp.f32 %f501, %f499, %f496, %p70;
|
|
selp.f32 %f502, %f500, %f498, %p70;
|
|
add.f32 %f503, %f501, 0fBF800000;
|
|
add.f32 %f489, %f501, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f488,%f489;
|
|
// inline asm
|
|
add.f32 %f504, %f503, %f503;
|
|
mul.f32 %f505, %f488, %f504;
|
|
mul.f32 %f506, %f505, %f505;
|
|
mov.f32 %f507, 0f3C4CAF63;
|
|
mov.f32 %f508, 0f3B18F0FE;
|
|
fma.rn.f32 %f509, %f508, %f506, %f507;
|
|
mov.f32 %f510, 0f3DAAAABD;
|
|
fma.rn.f32 %f511, %f509, %f506, %f510;
|
|
mul.rn.f32 %f512, %f511, %f506;
|
|
mul.rn.f32 %f513, %f512, %f505;
|
|
sub.f32 %f514, %f503, %f505;
|
|
neg.f32 %f515, %f505;
|
|
add.f32 %f516, %f514, %f514;
|
|
fma.rn.f32 %f517, %f515, %f503, %f516;
|
|
mul.rn.f32 %f518, %f488, %f517;
|
|
add.f32 %f519, %f513, %f505;
|
|
sub.f32 %f520, %f505, %f519;
|
|
add.f32 %f521, %f513, %f520;
|
|
add.f32 %f522, %f518, %f521;
|
|
add.f32 %f523, %f519, %f522;
|
|
sub.f32 %f524, %f519, %f523;
|
|
add.f32 %f525, %f522, %f524;
|
|
mov.f32 %f526, 0f3F317200;
|
|
mul.rn.f32 %f527, %f502, %f526;
|
|
mov.f32 %f528, 0f35BFBE8E;
|
|
mul.rn.f32 %f529, %f502, %f528;
|
|
add.f32 %f530, %f527, %f523;
|
|
sub.f32 %f531, %f527, %f530;
|
|
add.f32 %f532, %f523, %f531;
|
|
add.f32 %f533, %f525, %f532;
|
|
add.f32 %f534, %f529, %f533;
|
|
add.f32 %f535, %f530, %f534;
|
|
sub.f32 %f536, %f530, %f535;
|
|
add.f32 %f537, %f534, %f536;
|
|
mov.f32 %f538, 0f3EE8BA2E;
|
|
mul.rn.f32 %f539, %f538, %f535;
|
|
neg.f32 %f540, %f539;
|
|
fma.rn.f32 %f541, %f538, %f535, %f540;
|
|
fma.rn.f32 %f542, %f538, %f537, %f541;
|
|
mov.f32 %f543, 0f00000000;
|
|
fma.rn.f32 %f544, %f543, %f535, %f542;
|
|
add.rn.f32 %f545, %f539, %f544;
|
|
neg.f32 %f546, %f545;
|
|
add.rn.f32 %f547, %f539, %f546;
|
|
add.rn.f32 %f548, %f547, %f544;
|
|
mov.b32 %r259, %f545;
|
|
setp.eq.s32 %p71, %r259, 1118925336;
|
|
add.s32 %r260, %r259, -1;
|
|
mov.b32 %f549, %r260;
|
|
add.f32 %f550, %f548, 0f37000000;
|
|
selp.f32 %f551, %f549, %f545, %p71;
|
|
selp.f32 %f151, %f550, %f548, %p71;
|
|
mul.f32 %f552, %f551, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f553, %f552;
|
|
mov.f32 %f554, 0fBF317200;
|
|
fma.rn.f32 %f555, %f553, %f554, %f551;
|
|
mov.f32 %f556, 0fB5BFBE8E;
|
|
fma.rn.f32 %f557, %f553, %f556, %f555;
|
|
mul.f32 %f558, %f557, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f559, %f558;
|
|
add.f32 %f560, %f553, 0f00000000;
|
|
ex2.approx.f32 %f561, %f560;
|
|
mul.f32 %f562, %f559, %f561;
|
|
setp.lt.f32 %p72, %f551, 0fC2D20000;
|
|
selp.f32 %f563, 0f00000000, %f562, %p72;
|
|
setp.gt.f32 %p73, %f551, 0f42D20000;
|
|
selp.f32 %f976, 0f7F800000, %f563, %p73;
|
|
setp.eq.f32 %p74, %f976, 0f7F800000;
|
|
@%p74 bra BB0_78;
|
|
|
|
fma.rn.f32 %f976, %f976, %f151, %f976;
|
|
|
|
BB0_78:
|
|
setp.lt.f32 %p75, %f975, 0f00000000;
|
|
setp.eq.f32 %p76, %f149, 0f3F800000;
|
|
and.pred %p3, %p75, %p76;
|
|
mov.b32 %r261, %f976;
|
|
xor.b32 %r262, %r261, -2147483648;
|
|
mov.b32 %f564, %r262;
|
|
selp.f32 %f978, %f564, %f976, %p3;
|
|
setp.eq.f32 %p77, %f975, 0f00000000;
|
|
@%p77 bra BB0_81;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_81:
|
|
add.f32 %f567, %f975, %f975;
|
|
selp.f32 %f978, %f567, 0f00000000, %p76;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_146:
|
|
mov.u64 %rd210, image_HDR;
|
|
cvta.global.u64 %rd205, %rd210;
|
|
mov.u32 %r348, 8;
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd205, %r98, %r348, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f880, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f880;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs99, 0;
|
|
st.v4.u16 [%rd204], {%rs98, %rs98, %rs98, %rs99};
|
|
|
|
BB0_147:
|
|
ld.global.u32 %r349, [additive];
|
|
setp.eq.s32 %p144, %r349, 0;
|
|
@%p144 bra BB0_149;
|
|
|
|
mov.u64 %rd223, image_RNM0;
|
|
cvta.global.u64 %rd212, %rd223;
|
|
mov.u32 %r353, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r98, %r353, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs106, %rs107, %rs108, %rs109}, [%rd211];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f881, %rs106;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f882, %rs107;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f883, %rs108;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd217), _rt_buffer_get_64, (%rd212, %r98, %r353, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f884, %f881, 0f00000000;
|
|
add.f32 %f885, %f882, 0f00000000;
|
|
add.f32 %f886, %f883, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f886;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f885;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f884;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs110, 0;
|
|
st.v4.u16 [%rd217], {%rs103, %rs104, %rs105, %rs110};
|
|
bra.uni BB0_150;
|
|
|
|
BB0_149:
|
|
mov.u64 %rd230, image_RNM0;
|
|
cvta.global.u64 %rd225, %rd230;
|
|
mov.u32 %r355, 8;
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd225, %r98, %r355, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f887, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f887;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs112, 0;
|
|
st.v4.u16 [%rd224], {%rs111, %rs111, %rs111, %rs112};
|
|
|
|
BB0_150:
|
|
ld.global.u32 %r356, [additive];
|
|
setp.eq.s32 %p145, %r356, 0;
|
|
@%p145 bra BB0_152;
|
|
|
|
mov.u64 %rd243, image_RNM1;
|
|
cvta.global.u64 %rd232, %rd243;
|
|
mov.u32 %r360, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r98, %r360, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs119, %rs120, %rs121, %rs122}, [%rd231];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f888, %rs119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f889, %rs120;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f890, %rs121;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd237), _rt_buffer_get_64, (%rd232, %r98, %r360, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f891, %f888, 0f00000000;
|
|
add.f32 %f892, %f889, 0f00000000;
|
|
add.f32 %f893, %f890, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f893;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f892;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f891;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs123, 0;
|
|
st.v4.u16 [%rd237], {%rs116, %rs117, %rs118, %rs123};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd250, image_RNM1;
|
|
cvta.global.u64 %rd245, %rd250;
|
|
mov.u32 %r362, 8;
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd245, %r98, %r362, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f894, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f894;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs125, 0;
|
|
st.v4.u16 [%rd244], {%rs124, %rs124, %rs124, %rs125};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r363, [additive];
|
|
setp.eq.s32 %p146, %r363, 0;
|
|
@%p146 bra BB0_155;
|
|
|
|
mov.u64 %rd263, image_RNM2;
|
|
cvta.global.u64 %rd252, %rd263;
|
|
mov.u32 %r367, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r98, %r367, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs132, %rs133, %rs134, %rs135}, [%rd251];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f895, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f896, %rs133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f897, %rs134;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd257), _rt_buffer_get_64, (%rd252, %r98, %r367, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f898, %f895, 0f00000000;
|
|
add.f32 %f899, %f896, 0f00000000;
|
|
add.f32 %f900, %f897, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f900;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f899;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f898;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs136, 0;
|
|
st.v4.u16 [%rd257], {%rs129, %rs130, %rs131, %rs136};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd270, image_RNM2;
|
|
cvta.global.u64 %rd265, %rd270;
|
|
mov.u32 %r369, 8;
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd265, %r98, %r369, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f901, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f901;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs138, 0;
|
|
st.v4.u16 [%rd264], {%rs137, %rs137, %rs137, %rs138};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r370, [additive];
|
|
setp.eq.s32 %p147, %r370, 0;
|
|
@%p147 bra BB0_158;
|
|
|
|
mov.u64 %rd283, image_RNM3;
|
|
cvta.global.u64 %rd272, %rd283;
|
|
mov.u32 %r374, 8;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r98, %r374, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd271];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f902, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f903, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f904, %rs147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd277), _rt_buffer_get_64, (%rd272, %r98, %r374, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f905, %f902, 0f00000000;
|
|
add.f32 %f906, %f903, 0f00000000;
|
|
add.f32 %f907, %f904, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f907;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f906;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f905;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs149, 0;
|
|
st.v4.u16 [%rd277], {%rs142, %rs143, %rs144, %rs149};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd290, image_RNM3;
|
|
cvta.global.u64 %rd285, %rd290;
|
|
mov.u32 %r376, 8;
|
|
// inline asm
|
|
call (%rd284), _rt_buffer_get_64, (%rd285, %r98, %r376, %rd17, %rd18, %rd24, %rd24);
|
|
// inline asm
|
|
mov.f32 %f908, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f908;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs151, 0;
|
|
st.v4.u16 [%rd284], {%rs150, %rs150, %rs150, %rs151};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_79:
|
|
setp.geu.f32 %p78, %f975, 0f00000000;
|
|
@%p78 bra BB0_82;
|
|
|
|
cvt.rzi.f32.f32 %f566, %f538;
|
|
setp.neu.f32 %p79, %f566, 0f3EE8BA2E;
|
|
selp.f32 %f978, 0f7FFFFFFF, %f978, %p79;
|
|
|
|
BB0_82:
|
|
add.f32 %f568, %f150, 0f3EE8BA2E;
|
|
mov.b32 %r263, %f568;
|
|
setp.lt.s32 %p81, %r263, 2139095040;
|
|
@%p81 bra BB0_87;
|
|
|
|
setp.gtu.f32 %p82, %f150, 0f7F800000;
|
|
@%p82 bra BB0_86;
|
|
bra.uni BB0_84;
|
|
|
|
BB0_86:
|
|
add.f32 %f978, %f975, 0f3EE8BA2E;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_84:
|
|
setp.neu.f32 %p83, %f150, 0f7F800000;
|
|
@%p83 bra BB0_87;
|
|
|
|
selp.f32 %f978, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_87:
|
|
mul.f32 %f569, %f978, 0f437F0000;
|
|
setp.eq.f32 %p84, %f975, 0f3F800000;
|
|
selp.f32 %f570, 0f437F0000, %f569, %p84;
|
|
cvt.rzi.u32.f32 %r264, %f570;
|
|
cvt.u16.u32 %rs14, %r264;
|
|
mov.u16 %rs15, 255;
|
|
st.v2.u8 [%rd63], {%rs14, %rs15};
|
|
ld.global.u32 %r408, [imageEnabled];
|
|
|
|
BB0_88:
|
|
ld.global.f32 %f571, [lightColor];
|
|
mul.f32 %f162, %f970, %f571;
|
|
ld.global.f32 %f572, [lightColor+4];
|
|
mul.f32 %f163, %f970, %f572;
|
|
ld.global.f32 %f573, [lightColor+8];
|
|
mul.f32 %f164, %f970, %f573;
|
|
and.b32 %r265, %r408, 1;
|
|
setp.eq.b32 %p85, %r265, 1;
|
|
@!%p85 bra BB0_123;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_89:
|
|
mov.f32 %f576, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f577, %f576;
|
|
fma.rn.f32 %f578, %f577, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f165, %f578;
|
|
abs.f32 %f166, %f162;
|
|
setp.lt.f32 %p86, %f166, 0f00800000;
|
|
mul.f32 %f579, %f166, 0f4B800000;
|
|
selp.f32 %f580, 0fC3170000, 0fC2FE0000, %p86;
|
|
selp.f32 %f581, %f579, %f166, %p86;
|
|
mov.b32 %r266, %f581;
|
|
and.b32 %r267, %r266, 8388607;
|
|
or.b32 %r268, %r267, 1065353216;
|
|
mov.b32 %f582, %r268;
|
|
shr.u32 %r269, %r266, 23;
|
|
cvt.rn.f32.u32 %f583, %r269;
|
|
add.f32 %f584, %f580, %f583;
|
|
setp.gt.f32 %p87, %f582, 0f3FB504F3;
|
|
mul.f32 %f585, %f582, 0f3F000000;
|
|
add.f32 %f586, %f584, 0f3F800000;
|
|
selp.f32 %f587, %f585, %f582, %p87;
|
|
selp.f32 %f588, %f586, %f584, %p87;
|
|
add.f32 %f589, %f587, 0fBF800000;
|
|
add.f32 %f575, %f587, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f574,%f575;
|
|
// inline asm
|
|
add.f32 %f590, %f589, %f589;
|
|
mul.f32 %f591, %f574, %f590;
|
|
mul.f32 %f592, %f591, %f591;
|
|
mov.f32 %f593, 0f3C4CAF63;
|
|
mov.f32 %f594, 0f3B18F0FE;
|
|
fma.rn.f32 %f595, %f594, %f592, %f593;
|
|
mov.f32 %f596, 0f3DAAAABD;
|
|
fma.rn.f32 %f597, %f595, %f592, %f596;
|
|
mul.rn.f32 %f598, %f597, %f592;
|
|
mul.rn.f32 %f599, %f598, %f591;
|
|
sub.f32 %f600, %f589, %f591;
|
|
neg.f32 %f601, %f591;
|
|
add.f32 %f602, %f600, %f600;
|
|
fma.rn.f32 %f603, %f601, %f589, %f602;
|
|
mul.rn.f32 %f604, %f574, %f603;
|
|
add.f32 %f605, %f599, %f591;
|
|
sub.f32 %f606, %f591, %f605;
|
|
add.f32 %f607, %f599, %f606;
|
|
add.f32 %f608, %f604, %f607;
|
|
add.f32 %f609, %f605, %f608;
|
|
sub.f32 %f610, %f605, %f609;
|
|
add.f32 %f611, %f608, %f610;
|
|
mov.f32 %f612, 0f3F317200;
|
|
mul.rn.f32 %f613, %f588, %f612;
|
|
mov.f32 %f614, 0f35BFBE8E;
|
|
mul.rn.f32 %f615, %f588, %f614;
|
|
add.f32 %f616, %f613, %f609;
|
|
sub.f32 %f617, %f613, %f616;
|
|
add.f32 %f618, %f609, %f617;
|
|
add.f32 %f619, %f611, %f618;
|
|
add.f32 %f620, %f615, %f619;
|
|
add.f32 %f621, %f616, %f620;
|
|
sub.f32 %f622, %f616, %f621;
|
|
add.f32 %f623, %f620, %f622;
|
|
mov.f32 %f624, 0f3EE66666;
|
|
mul.rn.f32 %f625, %f624, %f621;
|
|
neg.f32 %f626, %f625;
|
|
fma.rn.f32 %f627, %f624, %f621, %f626;
|
|
fma.rn.f32 %f628, %f624, %f623, %f627;
|
|
mov.f32 %f629, 0f00000000;
|
|
fma.rn.f32 %f630, %f629, %f621, %f628;
|
|
add.rn.f32 %f631, %f625, %f630;
|
|
neg.f32 %f632, %f631;
|
|
add.rn.f32 %f633, %f625, %f632;
|
|
add.rn.f32 %f634, %f633, %f630;
|
|
mov.b32 %r270, %f631;
|
|
setp.eq.s32 %p88, %r270, 1118925336;
|
|
add.s32 %r271, %r270, -1;
|
|
mov.b32 %f635, %r271;
|
|
add.f32 %f636, %f634, 0f37000000;
|
|
selp.f32 %f637, %f635, %f631, %p88;
|
|
selp.f32 %f167, %f636, %f634, %p88;
|
|
mul.f32 %f638, %f637, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f639, %f638;
|
|
mov.f32 %f640, 0fBF317200;
|
|
fma.rn.f32 %f641, %f639, %f640, %f637;
|
|
mov.f32 %f642, 0fB5BFBE8E;
|
|
fma.rn.f32 %f643, %f639, %f642, %f641;
|
|
mul.f32 %f644, %f643, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f645, %f644;
|
|
add.f32 %f646, %f639, 0f00000000;
|
|
ex2.approx.f32 %f647, %f646;
|
|
mul.f32 %f648, %f645, %f647;
|
|
setp.lt.f32 %p89, %f637, 0fC2D20000;
|
|
selp.f32 %f649, 0f00000000, %f648, %p89;
|
|
setp.gt.f32 %p90, %f637, 0f42D20000;
|
|
selp.f32 %f979, 0f7F800000, %f649, %p90;
|
|
setp.eq.f32 %p91, %f979, 0f7F800000;
|
|
@%p91 bra BB0_91;
|
|
|
|
fma.rn.f32 %f979, %f979, %f167, %f979;
|
|
|
|
BB0_91:
|
|
setp.lt.f32 %p92, %f162, 0f00000000;
|
|
setp.eq.f32 %p93, %f165, 0f3F800000;
|
|
and.pred %p4, %p92, %p93;
|
|
mov.b32 %r272, %f979;
|
|
xor.b32 %r273, %r272, -2147483648;
|
|
mov.b32 %f650, %r273;
|
|
selp.f32 %f981, %f650, %f979, %p4;
|
|
setp.eq.f32 %p94, %f162, 0f00000000;
|
|
@%p94 bra BB0_94;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_94:
|
|
add.f32 %f653, %f162, %f162;
|
|
selp.f32 %f981, %f653, 0f00000000, %p93;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_92:
|
|
setp.geu.f32 %p95, %f162, 0f00000000;
|
|
@%p95 bra BB0_95;
|
|
|
|
cvt.rzi.f32.f32 %f652, %f624;
|
|
setp.neu.f32 %p96, %f652, 0f3EE66666;
|
|
selp.f32 %f981, 0f7FFFFFFF, %f981, %p96;
|
|
|
|
BB0_95:
|
|
add.f32 %f654, %f166, 0f3EE66666;
|
|
mov.b32 %r274, %f654;
|
|
setp.lt.s32 %p98, %r274, 2139095040;
|
|
@%p98 bra BB0_100;
|
|
|
|
setp.gtu.f32 %p99, %f166, 0f7F800000;
|
|
@%p99 bra BB0_99;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_99:
|
|
add.f32 %f981, %f162, 0f3EE66666;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_97:
|
|
setp.neu.f32 %p100, %f166, 0f7F800000;
|
|
@%p100 bra BB0_100;
|
|
|
|
selp.f32 %f981, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_100:
|
|
setp.eq.f32 %p101, %f162, 0f3F800000;
|
|
selp.f32 %f178, 0f3F800000, %f981, %p101;
|
|
abs.f32 %f179, %f163;
|
|
setp.lt.f32 %p102, %f179, 0f00800000;
|
|
mul.f32 %f657, %f179, 0f4B800000;
|
|
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p102;
|
|
selp.f32 %f659, %f657, %f179, %p102;
|
|
mov.b32 %r275, %f659;
|
|
and.b32 %r276, %r275, 8388607;
|
|
or.b32 %r277, %r276, 1065353216;
|
|
mov.b32 %f660, %r277;
|
|
shr.u32 %r278, %r275, 23;
|
|
cvt.rn.f32.u32 %f661, %r278;
|
|
add.f32 %f662, %f658, %f661;
|
|
setp.gt.f32 %p103, %f660, 0f3FB504F3;
|
|
mul.f32 %f663, %f660, 0f3F000000;
|
|
add.f32 %f664, %f662, 0f3F800000;
|
|
selp.f32 %f665, %f663, %f660, %p103;
|
|
selp.f32 %f666, %f664, %f662, %p103;
|
|
add.f32 %f667, %f665, 0fBF800000;
|
|
add.f32 %f656, %f665, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f655,%f656;
|
|
// inline asm
|
|
add.f32 %f668, %f667, %f667;
|
|
mul.f32 %f669, %f655, %f668;
|
|
mul.f32 %f670, %f669, %f669;
|
|
fma.rn.f32 %f673, %f594, %f670, %f593;
|
|
fma.rn.f32 %f675, %f673, %f670, %f596;
|
|
mul.rn.f32 %f676, %f675, %f670;
|
|
mul.rn.f32 %f677, %f676, %f669;
|
|
sub.f32 %f678, %f667, %f669;
|
|
neg.f32 %f679, %f669;
|
|
add.f32 %f680, %f678, %f678;
|
|
fma.rn.f32 %f681, %f679, %f667, %f680;
|
|
mul.rn.f32 %f682, %f655, %f681;
|
|
add.f32 %f683, %f677, %f669;
|
|
sub.f32 %f684, %f669, %f683;
|
|
add.f32 %f685, %f677, %f684;
|
|
add.f32 %f686, %f682, %f685;
|
|
add.f32 %f687, %f683, %f686;
|
|
sub.f32 %f688, %f683, %f687;
|
|
add.f32 %f689, %f686, %f688;
|
|
mul.rn.f32 %f691, %f666, %f612;
|
|
mul.rn.f32 %f693, %f666, %f614;
|
|
add.f32 %f694, %f691, %f687;
|
|
sub.f32 %f695, %f691, %f694;
|
|
add.f32 %f696, %f687, %f695;
|
|
add.f32 %f697, %f689, %f696;
|
|
add.f32 %f698, %f693, %f697;
|
|
add.f32 %f699, %f694, %f698;
|
|
sub.f32 %f700, %f694, %f699;
|
|
add.f32 %f701, %f698, %f700;
|
|
mul.rn.f32 %f703, %f624, %f699;
|
|
neg.f32 %f704, %f703;
|
|
fma.rn.f32 %f705, %f624, %f699, %f704;
|
|
fma.rn.f32 %f706, %f624, %f701, %f705;
|
|
fma.rn.f32 %f708, %f629, %f699, %f706;
|
|
add.rn.f32 %f709, %f703, %f708;
|
|
neg.f32 %f710, %f709;
|
|
add.rn.f32 %f711, %f703, %f710;
|
|
add.rn.f32 %f712, %f711, %f708;
|
|
mov.b32 %r279, %f709;
|
|
setp.eq.s32 %p104, %r279, 1118925336;
|
|
add.s32 %r280, %r279, -1;
|
|
mov.b32 %f713, %r280;
|
|
add.f32 %f714, %f712, 0f37000000;
|
|
selp.f32 %f715, %f713, %f709, %p104;
|
|
selp.f32 %f180, %f714, %f712, %p104;
|
|
mul.f32 %f716, %f715, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f717, %f716;
|
|
fma.rn.f32 %f719, %f717, %f640, %f715;
|
|
fma.rn.f32 %f721, %f717, %f642, %f719;
|
|
mul.f32 %f722, %f721, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f723, %f722;
|
|
add.f32 %f724, %f717, 0f00000000;
|
|
ex2.approx.f32 %f725, %f724;
|
|
mul.f32 %f726, %f723, %f725;
|
|
setp.lt.f32 %p105, %f715, 0fC2D20000;
|
|
selp.f32 %f727, 0f00000000, %f726, %p105;
|
|
setp.gt.f32 %p106, %f715, 0f42D20000;
|
|
selp.f32 %f982, 0f7F800000, %f727, %p106;
|
|
setp.eq.f32 %p107, %f982, 0f7F800000;
|
|
@%p107 bra BB0_102;
|
|
|
|
fma.rn.f32 %f982, %f982, %f180, %f982;
|
|
|
|
BB0_102:
|
|
setp.lt.f32 %p108, %f163, 0f00000000;
|
|
and.pred %p5, %p108, %p93;
|
|
mov.b32 %r281, %f982;
|
|
xor.b32 %r282, %r281, -2147483648;
|
|
mov.b32 %f728, %r282;
|
|
selp.f32 %f984, %f728, %f982, %p5;
|
|
setp.eq.f32 %p110, %f163, 0f00000000;
|
|
@%p110 bra BB0_105;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_105:
|
|
add.f32 %f731, %f163, %f163;
|
|
selp.f32 %f984, %f731, 0f00000000, %p93;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_103:
|
|
setp.geu.f32 %p111, %f163, 0f00000000;
|
|
@%p111 bra BB0_106;
|
|
|
|
cvt.rzi.f32.f32 %f730, %f624;
|
|
setp.neu.f32 %p112, %f730, 0f3EE66666;
|
|
selp.f32 %f984, 0f7FFFFFFF, %f984, %p112;
|
|
|
|
BB0_106:
|
|
add.f32 %f732, %f179, 0f3EE66666;
|
|
mov.b32 %r283, %f732;
|
|
setp.lt.s32 %p114, %r283, 2139095040;
|
|
@%p114 bra BB0_111;
|
|
|
|
setp.gtu.f32 %p115, %f179, 0f7F800000;
|
|
@%p115 bra BB0_110;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_110:
|
|
add.f32 %f984, %f163, 0f3EE66666;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_108:
|
|
setp.neu.f32 %p116, %f179, 0f7F800000;
|
|
@%p116 bra BB0_111;
|
|
|
|
selp.f32 %f984, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_111:
|
|
setp.eq.f32 %p117, %f163, 0f3F800000;
|
|
selp.f32 %f191, 0f3F800000, %f984, %p117;
|
|
abs.f32 %f192, %f164;
|
|
setp.lt.f32 %p118, %f192, 0f00800000;
|
|
mul.f32 %f735, %f192, 0f4B800000;
|
|
selp.f32 %f736, 0fC3170000, 0fC2FE0000, %p118;
|
|
selp.f32 %f737, %f735, %f192, %p118;
|
|
mov.b32 %r284, %f737;
|
|
and.b32 %r285, %r284, 8388607;
|
|
or.b32 %r286, %r285, 1065353216;
|
|
mov.b32 %f738, %r286;
|
|
shr.u32 %r287, %r284, 23;
|
|
cvt.rn.f32.u32 %f739, %r287;
|
|
add.f32 %f740, %f736, %f739;
|
|
setp.gt.f32 %p119, %f738, 0f3FB504F3;
|
|
mul.f32 %f741, %f738, 0f3F000000;
|
|
add.f32 %f742, %f740, 0f3F800000;
|
|
selp.f32 %f743, %f741, %f738, %p119;
|
|
selp.f32 %f744, %f742, %f740, %p119;
|
|
add.f32 %f745, %f743, 0fBF800000;
|
|
add.f32 %f734, %f743, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f733,%f734;
|
|
// inline asm
|
|
add.f32 %f746, %f745, %f745;
|
|
mul.f32 %f747, %f733, %f746;
|
|
mul.f32 %f748, %f747, %f747;
|
|
fma.rn.f32 %f751, %f594, %f748, %f593;
|
|
fma.rn.f32 %f753, %f751, %f748, %f596;
|
|
mul.rn.f32 %f754, %f753, %f748;
|
|
mul.rn.f32 %f755, %f754, %f747;
|
|
sub.f32 %f756, %f745, %f747;
|
|
neg.f32 %f757, %f747;
|
|
add.f32 %f758, %f756, %f756;
|
|
fma.rn.f32 %f759, %f757, %f745, %f758;
|
|
mul.rn.f32 %f760, %f733, %f759;
|
|
add.f32 %f761, %f755, %f747;
|
|
sub.f32 %f762, %f747, %f761;
|
|
add.f32 %f763, %f755, %f762;
|
|
add.f32 %f764, %f760, %f763;
|
|
add.f32 %f765, %f761, %f764;
|
|
sub.f32 %f766, %f761, %f765;
|
|
add.f32 %f767, %f764, %f766;
|
|
mul.rn.f32 %f769, %f744, %f612;
|
|
mul.rn.f32 %f771, %f744, %f614;
|
|
add.f32 %f772, %f769, %f765;
|
|
sub.f32 %f773, %f769, %f772;
|
|
add.f32 %f774, %f765, %f773;
|
|
add.f32 %f775, %f767, %f774;
|
|
add.f32 %f776, %f771, %f775;
|
|
add.f32 %f777, %f772, %f776;
|
|
sub.f32 %f778, %f772, %f777;
|
|
add.f32 %f779, %f776, %f778;
|
|
mul.rn.f32 %f781, %f624, %f777;
|
|
neg.f32 %f782, %f781;
|
|
fma.rn.f32 %f783, %f624, %f777, %f782;
|
|
fma.rn.f32 %f784, %f624, %f779, %f783;
|
|
fma.rn.f32 %f786, %f629, %f777, %f784;
|
|
add.rn.f32 %f787, %f781, %f786;
|
|
neg.f32 %f788, %f787;
|
|
add.rn.f32 %f789, %f781, %f788;
|
|
add.rn.f32 %f790, %f789, %f786;
|
|
mov.b32 %r288, %f787;
|
|
setp.eq.s32 %p120, %r288, 1118925336;
|
|
add.s32 %r289, %r288, -1;
|
|
mov.b32 %f791, %r289;
|
|
add.f32 %f792, %f790, 0f37000000;
|
|
selp.f32 %f793, %f791, %f787, %p120;
|
|
selp.f32 %f193, %f792, %f790, %p120;
|
|
mul.f32 %f794, %f793, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f795, %f794;
|
|
fma.rn.f32 %f797, %f795, %f640, %f793;
|
|
fma.rn.f32 %f799, %f795, %f642, %f797;
|
|
mul.f32 %f800, %f799, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f801, %f800;
|
|
add.f32 %f802, %f795, 0f00000000;
|
|
ex2.approx.f32 %f803, %f802;
|
|
mul.f32 %f804, %f801, %f803;
|
|
setp.lt.f32 %p121, %f793, 0fC2D20000;
|
|
selp.f32 %f805, 0f00000000, %f804, %p121;
|
|
setp.gt.f32 %p122, %f793, 0f42D20000;
|
|
selp.f32 %f985, 0f7F800000, %f805, %p122;
|
|
setp.eq.f32 %p123, %f985, 0f7F800000;
|
|
@%p123 bra BB0_113;
|
|
|
|
fma.rn.f32 %f985, %f985, %f193, %f985;
|
|
|
|
BB0_113:
|
|
setp.lt.f32 %p124, %f164, 0f00000000;
|
|
and.pred %p6, %p124, %p93;
|
|
mov.b32 %r290, %f985;
|
|
xor.b32 %r291, %r290, -2147483648;
|
|
mov.b32 %f806, %r291;
|
|
selp.f32 %f987, %f806, %f985, %p6;
|
|
setp.eq.f32 %p126, %f164, 0f00000000;
|
|
@%p126 bra BB0_116;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_116:
|
|
add.f32 %f809, %f164, %f164;
|
|
selp.f32 %f987, %f809, 0f00000000, %p93;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_114:
|
|
setp.geu.f32 %p127, %f164, 0f00000000;
|
|
@%p127 bra BB0_117;
|
|
|
|
cvt.rzi.f32.f32 %f808, %f624;
|
|
setp.neu.f32 %p128, %f808, 0f3EE66666;
|
|
selp.f32 %f987, 0f7FFFFFFF, %f987, %p128;
|
|
|
|
BB0_117:
|
|
add.f32 %f810, %f192, 0f3EE66666;
|
|
mov.b32 %r292, %f810;
|
|
setp.lt.s32 %p130, %r292, 2139095040;
|
|
@%p130 bra BB0_122;
|
|
|
|
setp.gtu.f32 %p131, %f192, 0f7F800000;
|
|
@%p131 bra BB0_121;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_121:
|
|
add.f32 %f987, %f164, 0f3EE66666;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_119:
|
|
setp.neu.f32 %p132, %f192, 0f7F800000;
|
|
@%p132 bra BB0_122;
|
|
|
|
selp.f32 %f987, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_122:
|
|
setp.eq.f32 %p133, %f164, 0f3F800000;
|
|
selp.f32 %f811, 0f3F800000, %f987, %p133;
|
|
cvt.u64.u32 %rd73, %r5;
|
|
cvt.u64.u32 %rd72, %r4;
|
|
mov.u64 %rd76, image;
|
|
cvta.global.u64 %rd71, %rd76;
|
|
// inline asm
|
|
call (%rd70), _rt_buffer_get_64, (%rd71, %r98, %r99, %rd72, %rd73, %rd24, %rd24);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f812, %f811;
|
|
mul.f32 %f813, %f812, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r295, %f813;
|
|
cvt.sat.f32.f32 %f814, %f191;
|
|
mul.f32 %f815, %f814, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r296, %f815;
|
|
cvt.sat.f32.f32 %f816, %f178;
|
|
mul.f32 %f817, %f816, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r297, %f817;
|
|
cvt.u16.u32 %rs16, %r295;
|
|
cvt.u16.u32 %rs17, %r297;
|
|
cvt.u16.u32 %rs18, %r296;
|
|
mov.u16 %rs19, 255;
|
|
st.v4.u8 [%rd70], {%rs16, %rs18, %rs17, %rs19};
|
|
ld.global.u32 %r408, [imageEnabled];
|
|
|
|
BB0_123:
|
|
cvt.u64.u32 %rd15, %r4;
|
|
cvt.u64.u32 %rd16, %r5;
|
|
and.b32 %r298, %r408, 4;
|
|
setp.eq.s32 %p134, %r298, 0;
|
|
@%p134 bra BB0_127;
|
|
|
|
ld.global.u32 %r299, [additive];
|
|
setp.eq.s32 %p135, %r299, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs20, %f392;}
|
|
|
|
// inline asm
|
|
@%p135 bra BB0_126;
|
|
|
|
mov.u64 %rd89, image_HDR;
|
|
cvta.global.u64 %rd78, %rd89;
|
|
mov.u32 %r303, 8;
|
|
// inline asm
|
|
call (%rd77), _rt_buffer_get_64, (%rd78, %r98, %r303, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs27, %rs28, %rs29, %rs30}, [%rd77];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f819, %rs27;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f820, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f821, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd83), _rt_buffer_get_64, (%rd78, %r98, %r303, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f822, %f162, %f819;
|
|
add.f32 %f823, %f163, %f820;
|
|
add.f32 %f824, %f164, %f821;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f824;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f823;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs24, %f822;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd83], {%rs24, %rs25, %rs26, %rs20};
|
|
bra.uni BB0_127;
|
|
|
|
BB0_126:
|
|
mov.u64 %rd96, image_HDR;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
mov.u32 %r305, 8;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r98, %r305, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f164;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f163;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f162;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd90], {%rs31, %rs32, %rs33, %rs20};
|
|
|
|
BB0_127:
|
|
selp.f32 %f829, 0f3F800000, 0f3E800000, %p148;
|
|
mul.f32 %f204, %f829, %f971;
|
|
mul.f32 %f205, %f829, %f972;
|
|
mul.f32 %f206, %f829, %f973;
|
|
mul.f32 %f207, %f829, %f974;
|
|
ld.global.f32 %f830, [lightColor];
|
|
mul.f32 %f208, %f204, %f830;
|
|
ld.global.f32 %f831, [lightColor+4];
|
|
mul.f32 %f209, %f204, %f831;
|
|
ld.global.f32 %f832, [lightColor+8];
|
|
mul.f32 %f210, %f204, %f832;
|
|
ld.global.u32 %r306, [additive];
|
|
setp.eq.s32 %p136, %r306, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f392;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd109, image_RNM0;
|
|
cvta.global.u64 %rd98, %rd109;
|
|
mov.u32 %r310, 8;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r98, %r310, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs41, %rs42, %rs43, %rs44}, [%rd97];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f833, %rs41;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f834, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f835, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd103), _rt_buffer_get_64, (%rd98, %r98, %r310, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f836, %f208, %f833;
|
|
add.f32 %f837, %f209, %f834;
|
|
add.f32 %f838, %f210, %f835;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f838;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f837;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f836;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd103], {%rs38, %rs39, %rs40, %rs34};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd116, image_RNM0;
|
|
cvta.global.u64 %rd111, %rd116;
|
|
mov.u32 %r312, 8;
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd111, %r98, %r312, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f210;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f209;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f208;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs45, %rs46, %rs47, %rs34};
|
|
|
|
BB0_130:
|
|
mov.f32 %f843, 0f34000000;
|
|
max.f32 %f844, %f204, %f843;
|
|
div.rn.f32 %f845, %f205, %f844;
|
|
fma.rn.f32 %f211, %f845, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f846, %f206, %f844;
|
|
fma.rn.f32 %f212, %f846, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f847, %f207, %f844;
|
|
fma.rn.f32 %f213, %f847, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r313, [additive];
|
|
setp.eq.s32 %p137, %r313, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f392;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_132;
|
|
|
|
mov.u64 %rd129, image_RNM1;
|
|
cvta.global.u64 %rd118, %rd129;
|
|
mov.u32 %r317, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r98, %r317, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs55, %rs56, %rs57, %rs58}, [%rd117];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f848, %rs55;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f850, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd123), _rt_buffer_get_64, (%rd118, %r98, %r317, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f851, %f211, %f848;
|
|
add.f32 %f852, %f211, %f849;
|
|
add.f32 %f853, %f211, %f850;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f853;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f852;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f851;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd123], {%rs52, %rs53, %rs54, %rs48};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd136, image_RNM1;
|
|
cvta.global.u64 %rd131, %rd136;
|
|
mov.u32 %r319, 8;
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd131, %r98, %r319, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f211;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs59, %rs59, %rs59, %rs48};
|
|
|
|
BB0_133:
|
|
ld.global.u32 %r320, [additive];
|
|
setp.eq.s32 %p138, %r320, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f392;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_135;
|
|
|
|
mov.u64 %rd149, image_RNM2;
|
|
cvta.global.u64 %rd138, %rd149;
|
|
mov.u32 %r324, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r98, %r324, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs67, %rs68, %rs69, %rs70}, [%rd137];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f856, %rs67;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f857, %rs68;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs69;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd143), _rt_buffer_get_64, (%rd138, %r98, %r324, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f859, %f212, %f856;
|
|
add.f32 %f860, %f212, %f857;
|
|
add.f32 %f861, %f212, %f858;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f861;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f860;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f859;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd143], {%rs64, %rs65, %rs66, %rs60};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd156, image_RNM2;
|
|
cvta.global.u64 %rd151, %rd156;
|
|
mov.u32 %r326, 8;
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd151, %r98, %r326, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f212;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs71, %rs71, %rs71, %rs60};
|
|
|
|
BB0_136:
|
|
ld.global.u32 %r327, [additive];
|
|
setp.eq.s32 %p139, %r327, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f392;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_138;
|
|
|
|
mov.u64 %rd169, image_RNM3;
|
|
cvta.global.u64 %rd158, %rd169;
|
|
mov.u32 %r331, 8;
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd158, %r98, %r331, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
ld.v4.u16 {%rs79, %rs80, %rs81, %rs82}, [%rd157];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f864, %rs79;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f865, %rs80;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f866, %rs81;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd163), _rt_buffer_get_64, (%rd158, %r98, %r331, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
add.f32 %f867, %f213, %f864;
|
|
add.f32 %f868, %f213, %f865;
|
|
add.f32 %f869, %f213, %f866;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f869;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f868;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f867;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd163], {%rs76, %rs77, %rs78, %rs72};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd176, image_RNM3;
|
|
cvta.global.u64 %rd171, %rd176;
|
|
mov.u32 %r333, 8;
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd171, %r98, %r333, %rd15, %rd16, %rd24, %rd24);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f213;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd170], {%rs83, %rs83, %rs83, %rs72};
|
|
|
|
BB0_159:
|
|
ret;
|
|
}
|
|
|
|
|